Changelog
Every change to TopxAI's catalogue, routes, prices and console, dated: new models, new endpoints, billing rules and fixes. RSS feed.
Exact numbers in converted tool calls, and a few relay and console fixes
· fixed
On /v1/messages, tool call arguments from GPT, Grok, Kimi, GLM and DeepSeek keep large numbers exactly; stop lists, streams and the console are fixed too.
Corrected on November 1, 2026: the stream item now says what each endpoint does with an unusual [DONE] line, and the list of cases that still round is complete.
A relay release on November 1 fixes how some requests are converted, plus two issues in the console. No model or price changed, and nothing needs to change on your side.
Conversion:
On /v1/messages, the arguments of a tool call in a non-streamed answer from a GPT, Grok, Kimi, GLM or DeepSeek model keep their numbers exactly as the model wrote them. Integers above 2^53 and long decimals used to be rounded, so 9007199254740993 arrived as 9007199254740992. This includes GPT-6.1 Sol and Terra, which answer through the Responses API. Streamed answers already passed the arguments through unchanged.
On /v1/chat/completions, a stop list for a Claude model that holds a non-string item, such as ["END", 1], no longer fails with a 500. The string items are sent as stop sequences and the rest is ignored.
A bare [DONE] line from an upstream, without data:, now ends the stream cleanly; /v1/chat/completions used to send a broken data: ] event first. A line that only starts with [DONE] no longer ends the stream: /v1/messages now delivers the rest of the answer, and /v1/chat/completions passes the line on unchanged.
Still rounded: tool input in a non-streamed Claude answer on /v1/chat/completions; tool calls in the history of a /v1/messages request, whichever model it goes to; and tool calls in the history of a /v1/chat/completions request to a Claude model. If such a number must stay exact, send it as a string.
Console and website:
Chart tooltips show key names and usernames as plain text. The chart library used to insert a tooltip title as HTML, so a name containing HTML tags was rendered as markup.
axios and DOMPurify, which the website and console use, are updated to versions with no published security advisories.
Signing out no longer shows "Session expired" next to "Signed out".
Answers no longer carry billing_usage, and Astra drops log probabilities on Responses
· fixed
Converted answers stop showing the internal usage.billing_usage object, and gpt-6-astra on /v1/responses drops top_logprobs instead of failing with a 400.
A relay release on October 30 fixes two things you could see in answers. No model or price changed.
When the endpoint you call and the protocol of the model's channel differ, the relay converts the answer: Chat Completions or Messages from GPT-6.1 Sol and Terra, which OpenAI serves through Responses; Messages from a Chat Completions channel such as Grok; Chat Completions or Responses from Claude. Those answers used to include usage.billing_usage, an internal object holding the channel's original token counts that TopxAI bills on. It no longer appears, in a whole answer or in a stream event. Billing still reads the same counts, so charges do not change. If your code read that object, use the usage fields your endpoint defines; they report the same usage in that endpoint's format.
gpt-6-astra on /v1/responses with top_logprobs was rejected by the upstream with a 400. TopxAI now removes top_logprobs and the message.output_text.logprobs item of include before forwarding, as it already does for gpt-6.1-terra, and the answer comes back without log probabilities. Other include items, the reasoning effort, temperature, top_p and the output cap are sent as before.
The four model IDs replaced earlier in October (claude-sonnet-5, jev-1.13.0, kimi-k3 and deepseek-flash) keep returning HTTP 403. Nothing changes for callers: the relay now treats them like other retired IDs and removes them from channels and price tables at startup.
Charges follow the usage your answer shows
· fixed
A usage.billing_usage object that an upstream adds to an answer is now ignored and removed; requests are billed on the usage the answer reports.
Some upstream gateways add their own usage.billing_usage object to answers they convert. From a second relay release on October 30, TopxAI ignores that object: a request is billed on the usage fields your endpoint returns, and an answer TopxAI converts itself is still billed on the channel's original token counts. The object is also removed before the answer reaches you, on endpoints TopxAI passes through unchanged as well. Prices did not change.
GPT-6.1 Terra joins the OpenAI routes
· new, fixed
New model gpt-6.1-terra: $0.50 input and $2.50 output per million tokens on the shared route, with reasoning that can be turned off.
The OpenAI shared and official routes, and Auto route, now serve gpt-6.1-terra next to gpt-6.1-sol and gpt-6-astra. No model is retired. Route keys without a model restriction can call it with the same key; a key with an allowed-model list needs gpt-6.1-terra added to that list.
Default USD prices per million tokens, based on OpenAI's model page, checked October 30, 2026:
Category Shared (50%) Official (90%) OpenAI Standard
Input $0.50 $0.90 $1
Output $2.50 $4.50 $5
Cache read $0.025 $0.045 $0.05
Cache write $0.625 $1.125 $1.25
Above 272,000 input tokens, including cache tokens, the whole request uses the long-context tier: input and cache rates double, output rates multiply by 1.5. Cache writes are sold at OpenAI's default 30-minute retention.
The context window is 1,050,000 tokens, of which up to 922,000 can be input, with up to 128,000 output tokens. Input is text or images; output is text. Reasoning levels are none, low, medium (default), high, xhigh and max. Unlike Sol, Terra can turn reasoning off: none, and Anthropic-style disabled thinking, reach OpenAI as none, and only then are temperature and top_p forwarded. Terra returns no logprobs at any level: the upstream rejects top_logprobs and the logprobs include, so TopxAI removes them. Legacy minimal becomes low. Chat Completions and Anthropic-compatible requests use Responses upstream, because OpenAI serves this model's tool calls only there; the client's response format and tools are kept.
Fixed for every OpenAI model: a 12-hour ttl set on a single prompt_cache_breakpoint inside input was forwarded, so a write the provider prices at twice the default could be billed at the default cache-write price. It is now dropped, like the request-level ttl in prompt_cache_options, and the breakpoint keeps the default lifetime.
See the launch post and current prices.
MCP ask_model: room to reason, a reasoning_effort argument, and errors for empty replies
· new, improved, fixed
ask_model defaults to 8,192 output tokens and takes a reasoning_effort argument. An empty reply is now an error that reports its charge.
Two relay releases on October 30 change the ask_model tool of the TopxAI Models MCP server (/mcp/models). Prices and billing rules are unchanged. MCP clients read tool definitions when they connect, so reconnect the server or start a new session to see the new argument.
Reasoning:
Every text model in the catalogue reasons before it answers, and the reasoning counts toward max_output_tokens without being returned. The default is now 8,192 tokens instead of 2,048, and the tool description asks for at least 4,096. With the old default a model could spend the whole budget thinking and write nothing.
ask_model takes an optional reasoning_effort. Each model takes its own set, and list_models shows the efforts and the default of each one: claude-opus-5-5, for example, takes low through max, GLM-5.3-Abliterated takes low, high and max, and deepseek-flash-Abliterated can also turn reasoning off with none. The default applies when you leave the argument out. An effort the model does not take is refused before anything is reserved or charged.
max_cost_usd is checked against the whole max_output_tokens, so with the new default the estimate covers 8,192 output tokens. For a tight budget, pass a smaller max_output_tokens. The charge still counts only the tokens the model writes.
Empty replies:
When a model returns no text, ask_model now answers with isError: true and names the cause: the limit ran out while the model was still reasoning (raise max_output_tokens or lower reasoning_effort); the model declined, in which case the refusal is quoted and the same prompt sent again is likely to be declined and charged again; or the model stopped without writing anything.
The provider still ran the request, so it is charged like the same API call plus the call fee, and structuredContent keeps the usage and the exact charge. Before this release such a reply came back as a success, with a text block that lacked its text member.
Claude Opus 5.5 special offer: a low-price line that is often unavailable
· new
A temporary Opus 5.5 line at $0.45 input and $0.90 output per million tokens. It is often down, sometimes for long stretches, with no availability guarantee.
Read this first: the special-offer line is often unavailable, sometimes for long stretches, and its availability is not guaranteed. On its first day the upstream stopped serving six minutes after the line opened and was still failing every request more than half an hour later. Expect that to happen again. While it is down, requests get HTTP 503 overloaded_error and are not charged.
claude-opus-5-5 now has a separate, temporary route, Special offer (claude-special), at a much lower fixed price. USD per 1M tokens:
Lane Special offer Shared pool Anthropic list price
Input $0.45 $2.00 $4.00
Output $0.90 $10.00 $20.00
Cache read $0.10 $0.10 $0.20
Cache write (5 min) $0.20 $2.50 $5.00
Cache write (1 h) $0.30 $4.00 $8.00
Only a key created on this line uses it. Choose Special offer under Claude when you create a key, or use Create a key on this line on the home-page card. Existing keys and Auto route keys never switch to it.
It sells Opus 5.5 only; any other model on such a key gets HTTP 403 before anything is charged.
There is no fallback. A failed request is not retried on another route, and keys on the line may stop working. Keep a key on the shared pool or the official line for work that has to go through.
The line adds its own system prompt to every request. Those tokens are billed like your own, so check your usage log.
We guarantee that requests are answered by the official Claude Opus 5.5 model. Speed, quality, feature support and an experience like the official API are not guaranteed.
The offer is temporary and can end at short notice.
All terms, the extra-token measurements and Claude Code settings are on the special-offer page.
DeepSeek V4 Pro retired; Abliterated models grouped together
· maintenance
deepseek-v4-pro is no longer offered. The two private Abliterated models now share a display group, with their routes and prices unchanged.
deepseek-v4-pro has left the catalogue. New requests for it return HTTP 403 with the public error code permission_denied. It no longer appears in the model list or price table, while its standalone model reference remains available, clearly marked as currently unavailable. The page keeps its historical overview and links to other models; it does not offer a current price or API access. No alias forwards requests to another model.
Past usage records and captured charges remain intact. If a key permits only this model, edit its allowed models before choosing another available model; its permissions are not expanded automatically.
GLM-5.3-Abliterated and deepseek-flash-Abliterated now appear together under Abliterated. Both are private deployments. Their model IDs, routes, prices and capabilities are unchanged: GLM remains on z-official, and Flash remains on deepseek-official.
deepseek-flash-Abliterated replaces the old Flash model
· improved
A private Flash deployment with image inputs, fixed token prices and a 262,144-token output limit. Update saved model IDs and restricted keys.
Update, 2026-10-25: deepseek-v4-pro is retired. New requests for that ID return HTTP 403 permission_denied; the launch details below are historical.
deepseek-flash-Abliterated replaces deepseek-flash on the existing deepseek-official route. It uses a private deployment, accepts text and images, has a 1,048,576-token context window and supports up to 262,144 output tokens. Chat Completions, stateless Responses and the Messages bridge are supported.
Default prices per million tokens are $1.20 input, $0.15 cache read and $2.40 output, with no peak/off-peak tier or long-context premium. deepseek-v4-pro retains its official deployment and 80%-of-list peak/off-peak prices. Accepted requests and historical charges retain their captured prices.
Reasoning effort accepts none, low, high and max, with high as the default. The old deepseek-flash ID returns HTTP 403 and has no alias. Update saved model IDs, output limits and the allowed-model lists of restricted keys. Unrestricted keys and the base URL can stay the same.
Because fine-tuning may affect safety alignment, the model may produce biased, offensive or locally unlawful content. Users bear full responsibility for compliance and legal consequences. The Acceptable Use Policy applies. See the migration post for the client changes.
Kimi K3.1 Pro replaces Kimi K3
· improved
The model ID changes to kimi-k3.1-pro. Default TopxAI token prices stay unchanged and equal 120% of the new official reference.
The kimi-official route now serves kimi-k3.1-pro. Replace kimi-k3 in client settings; the retired ID returns HTTP 403 without a forwarding alias. Keys restricted to specific models need the new ID added explicitly. Unrestricted route keys and the base URL are unchanged.
Default prices per million tokens remain $2.40 input, $0.24 cache read, $2.40 default 5-minute cache write and $12 output. The new official reference, checked October 24, 2026, is $2 / $0.20 / $2 / $10 in the same order. TopxAI is therefore 120% of that reference, or 20% higher. No long-context tier is added. Administrator overrides, deliberate deletions and historical charges are preserved.
The context window is 1,048,576 tokens. TopxAI keeps its published output limit at 131,072 tokens to leave room for the prompt. Thinking stays enabled; use reasoning_effort: "low", "high" or "max" (default). See the migration guide and request example.
Claude Sonnet 5.5 cache reads at 50% and 90% of list
· fixed
Sonnet 5.5 cache reads now cost $0.05 per million tokens on the shared route and $0.09 on the official route, half of what TopxAI charged since October 2.
Anthropic prices Claude Sonnet 5.5 cache reads at 0.05x the input price, $0.10 per million tokens, which is half of Sonnet 5's rate. When Sonnet 5.5 replaced Sonnet 5 on October 2, TopxAI's price table kept Sonnet 5's $0.20 as the list price, so cache reads cost $0.10 on the shared route and $0.18 on the official route instead of 50% and 90% of list. The table now matches Anthropic. Input, output and cache-write prices do not change, and requests accepted before the change settle at the price they were quoted.
Claude Sonnet 5.5 cache read, USD per million tokens:
Route Until October 23 Now Share of the $0.10 list price
Shared $0.10 $0.05 50%
Official $0.18 $0.09 90%
Rates follow Anthropic's pricing page and Sonnet 5.5 model page, checked October 23, 2026. Opus 5.5 and Fable 5.1 cache reads were already at 50% and 90% of their list prices. The model page lists every lane, and the usage log shows cache-read tokens separately from other input.
Images in tool results, fairer image and cache billing, and a round of relay fixes
· improved, fixed
Claude Code image reads no longer fail on Kimi, DeepSeek, GLM and Grok, an image answer without an image is free, plus conversion, streaming and security fixes.
October 24, 2026 update: kimi-k3 has been replaced by Kimi K3.1 Pro. Its default TopxAI token prices are unchanged and now equal 120% of the new model’s official reference. Update the model ID and any restricted key allowlist; the old ID returns HTTP 403. The original release record below is preserved.
Two relay releases on October 23 change how some requests are converted and billed. No model or price changed, and nothing needs to change on your side. The blog post explains the larger items.
Requests that used to fail:
On /v1/messages, an image inside a tool_result (a screenshot, an image file Claude Code read) no longer fails with a 400 on Kimi, DeepSeek, GLM and Grok. Its text stays in the tool message and the image moves to the user message that follows. A document inside a tool_result is still a 400 on those models.
GLM-5.3-Abliterated now gets a note in place of each image or document on /v1/messages too, as it already did on /v1/chat/completions.
Two consecutive assistant messages with tool calls no longer lose the first message's calls when sent to Claude, which Claude answered with a 400.
Responses input that ends with the model's own answer (a client retrying a stream) is no longer sent to Claude as a prefill, which Claude 5.x rejects.
A Responses tool output made of content parts (text and images) reaches Claude as text and image blocks.
Kimi K3 messages that declare tools mid-conversation keep their tools.
Per-message output_config and safeguards reach Claude on /v1/messages, and every anthropic-beta header line is forwarded, not just the first.
Billing:
A Claude stream whose prompt came entirely from the cache no longer gets an estimate of the prompt added when it ends early.
Images are counted by what came back: an entry with only revised_prompt is not billed, and one image sent as two entries is billed once. An answer with no image at all is an error and nothing is charged.
The cache TTL in prompt_cache_options is not forwarded to OpenAI or Kimi, so every cache write is the default kind at the listed cache-write price.
A Responses stream that fails after you received output is settled on the usage the failure event reports, when it reports output.
Streams and limits:
Reasoning items in converted Responses streams always carry summary, and answer text that arrives in the same chunk as reasoning is no longer dropped on /v1/messages.
A slow connection no longer loses a large image event halfway: each stream gets time in proportion to what it has been sent.
gpt-6.1-sol keeps the order of your JSON Schema properties on Chat Completions.
Long Claude Code sessions on Kimi, DeepSeek, GLM and Grok convert much faster: a 400-turn conversation went from about one second to 21 ms.
A request body over the size limit answers 413 instead of 400.
Security:
/api/usage/token/ and /api/log/token now apply the key's IP allow list, as /v1 does.
The relay runs on Go 1.26.10, with the standard library's TLS, certificate and HTTP fixes since 1.26.6, a redirect from https to plain http no longer carries credentials, and a redirect to another host no longer carries the provider key header.
Jev 1.14 replaces Jev 1.13, with images and Box questions
· new, improved
Use jev-1.14.0 on /v1/systemone at the same price. Questions can now look at PNG and JPEG images, and the new Box type finds where something is in a picture.
jev-1.14.0 is now the Jev model on POST /v1/systemone. The price is unchanged: $0.04494 per 1M input tokens, 107% of TypeSafe's list price of $0.042 checked on its models page on October 16, 2026, and output stays free. The old jev-1.13.0 ID is retired without a forwarding alias; a request for it gets a 403. jev-latest and jev-preview keep working and are answered by jev-1.14.0, so every response now reports "model": "jev-1.14.0". If you tuned thresholds on 1.13, check them again. Unrestricted keys keep working; if a key has an allowed-model list that names jev-1.13.0, add jev-1.14.0 to it, which also covers jev-latest and jev-preview.
Questions can now look at images. List each PNG or JPEG once in a top-level assets map, base64-encoded or as a file part of a multipart/form-data request, and write its reference string in the state, an instruction or a criterion. The new box question type returns where a target is in an image as four edges (top, bottom, left, right, fractions of the image) with a confidence. The latest TypeSafe SDKs (Python 0.7.5, JavaScript 0.7.0) upload images as multipart and compress request bodies with zstd; both work against TopxAI's base URL.
TypeSafe labels images and Box experimental, and its API requires the TypeSafe-Danger-Zone-Images: true header for them. TopxAI adds the header whenever a request carries an image or a Box question. Image tokens are billed as input tokens at the same price and are included in usage.input_tokens: in our October 16 measurements an image cost ceil(width / 28) × ceil(height / 28) × 2 tokens per use, so 1,276 tokens for an 800 × 600 photo.
A System One request may now be up to 32 MiB, with state and questions together still at most 1 MiB. When TypeSafe rejects a request with a 400 or 413 (for example an image with a side under 280 pixels, which TypeSafe rejects although its docs say 256), its message reaches you as written and nothing is charged, the same as its 422 detail before. The Jev documentation covers limits, multipart uploads and Box, and the launch post has our measurements of Box accuracy and image token counts.
Claude Sonnet 5.5 replaces Sonnet 5 at the same prices
· improved
Use claude-sonnet-5-5 on both Claude routes. Input, output and cache prices stay the same; update saved model IDs and restricted keys.
Select claude-sonnet-5-5 on the Claude shared or official route, or with an Auto-route key. The old claude-sonnet-5 ID is retired and has no forwarding alias. Unrestricted keys can keep using the same key; if a key has an allowed-model list, edit that list to include the new ID.
Default prices in USD per million tokens, checked against Anthropic's Sonnet 5.5 model page on October 2, 2026:
Category Shared (50%) Official (90%) Anthropic list
Input $1 $1.80 $2
Output $5 $9 $10
Cache read $0.10 $0.18 $0.20
Cache write, 5 minutes $1.25 $2.25 $2.50
Cache write, 1 hour $2 $3.60 $4
All five rates remain unchanged from Sonnet 5. (Correction, October 23, 2026: Anthropic lists Sonnet 5.5 cache reads at $0.10, not $0.20, so the cache-read prices above were twice the 50% and 90% shown; they are now $0.05 and $0.09, see the correction.) The model supports a 1M-token context and up to 128K output tokens, with no long-context price premium. Accepted requests and historical charges keep their captured prices.
Adaptive thinking defaults to high effort. TopxAI maps legacy disabled-thinking and OpenAI none requests to between_tools at low effort, and converts manual thinking budgets to effort. Unsupported sampling parameters are removed. Forced tool selection returns a 400 error without a charge; use automatic tool selection. Native thinking signatures and tool-result conversations are preserved.
See the migration post for supported settings, and the Claude Code guide for the model variables to update. Provider-native computer-use clients need Anthropic's new toolset format, described in its migration guide.
During the October 2 supplier checks, some between_tools requests returned 502, including medium on both routes and OpenAI none on the official route. Adaptive mode with explicit low effort passed on both routes: use thinking: {"type":"adaptive"} with output_config: {"effort":"low"} for Messages, or reasoning_effort: "low" for Chat Completions, when it fits your task. Adaptive mode can still think before answering, and those tokens are billed as output. TopxAI preserves the requested mode rather than silently substituting another one.
GPT-6.1 Sol replaces GPT-6 Sol
· improved
Upgrade to gpt-6.1-sol: shared routes remain $1 input and $5 output per million tokens, with cache reads reduced to $0.05.
The OpenAI shared and official routes move to gpt-6.1-sol. Update the model ID in your client: gpt-6-sol is retired, including its reasoning suffixes, with no forwarding alias. For keys restricted to specific models, update the allowed-model list too. Unrestricted keys and route keys without a model restriction keep working with the same key.
Default USD prices per million tokens, based on OpenAI's model page, checked September 30, 2026:
Category Shared (50%) Official (90%) OpenAI Standard
Input $1 $1.80 $2
Output $5 $9 $10
Cache read $0.05 $0.09 $0.10
Cache write $1.25 $2.25 $2.50
Cache reads cost half as much as on GPT-6 Sol; the other standard rates are unchanged. Above 272,000 input tokens, including cache tokens, the whole request uses the long-context tier: input and cache rates double, output rates multiply by 1.5. Existing administrator route-price overrides are preserved; accepted requests and historical charges retain their captured prices.
The context window remains 1,050,000 tokens, with up to 128,000 output tokens. Reasoning levels are low, medium (default), high, xhigh and max. Reasoning cannot be disabled; legacy none and minimal are normalized to low. Chat Completions and Anthropic-compatible requests use Responses upstream, preserving the client's response format and tools.
The Codex installer uses the new model by default. See the upgrade guide and current prices.
GLM-5.3-Abliterated publishes its reasoning modes and 1M-token context
· improved, fixed
The price feed now lists the low, high and max reasoning modes and the 1,000,000-token window, and the Z.ai-style thinking switch no longer fails with a 400.
GLM-5.3-Abliterated runs on a deployment with a 1,000,000-token context and three reasoning modes. /api/pricing now publishes both: context_window 1000000, max_output 131072, and the reasoning efforts low, high and max, with max as the default. TopxAI Desktop reads the feed, so it now offers the effort picker for this model and plans for the full window. Until now its requests always ran at max and it treated the model as a 128K-token window.
Fixed: switching reasoning the way Z.ai's own API does, with "thinking": {"type": "disabled"} or {"type": "enabled"}, failed with a 400, because this deployment rejects the object. The relay now turns disabled into reasoning_effort: "none", which runs as low with the reasoning text left out, and drops enabled. The boolean form "thinking": false is accepted as well; on every model it used to fail with a JSON error, and other lines now receive it as the object form.
Prices do not change. The model page explains where the reasoning text appears and how its tokens are billed.
Claude Fable 5 is retired; use Fable 5.1
· maintenance
claude-fable-5 is no longer offered. Claude Fable 5.1 stays at the same input and output prices with a lower cache-read price.
claude-fable-5 has left the catalogue. New requests for it are refused with HTTP 403 and the error code model_not_offered, and it no longer appears in /v1/models or the price table. Switch to claude-fable-5-1; it is not selected automatically.
Fable 5.1 costs the same per input and output token on both Claude routes ($5 / $25 shared, $9 / $45 official, USD per million tokens), and its cache reads are cheaper: $0.125 shared and $0.225 official, against $0.50 and $0.90 for Fable 5. Cache writes are unchanged.
Past requests to Fable 5 keep their usage-log rows and the prices they were charged. The catalogue now holds 14 models on 7 lines.
Old model IDs claude-opus-5, gpt-5.6-sol and grok-4.6 are retired
· maintenance
The three old IDs no longer forward to their successors. Use claude-opus-5-5, gpt-6-sol and grok-4.7; prices do not change.
claude-opus-5, gpt-5.6-sol and grok-4.6 stopped working on September 24. Until now each of them forwarded to its upgraded model. Requests for them are now refused with HTTP 403 and the error code model_not_offered, and they no longer appear in /v1/models or the price table.
Switch to the new IDs: claude-opus-5-5, gpt-6-sol and grok-4.7. Their prices are in the upgrade notes for Opus 5.5, GPT-6 Sol and Grok 4.7; this change does not touch them.
An API key whose model limits name only an old ID can no longer reach the new model. Edit the key's model limits and select the new ID.
Past requests made with the old IDs keep their usage-log rows and the prices they were charged.
Claude Opus 5.5 replaces Opus 5 with lower token prices
· improved
Opus 5.5 on both Claude routes at 50% and 90% of the Anthropic list price, including cache reads and writes. Existing Opus 5 API calls remain compatible.
Select claude-opus-5-5 on the shared or official Claude route. Existing calls to claude-opus-5, including reasoning-effort suffixes, use Opus 5.5 at the new prices. Existing API keys restricted to Opus 5 can use the new model.
Prices in USD per million tokens, based on Anthropic's published standard prices, verified September 23, 2026:
Category Shared (50%) Official (90%) Anthropic list
Input $2 $3.60 $4
Output $10 $18 $20
Cache read $0.10 $0.18 $0.20
Cache write, 5 minutes $2.50 $4.50 $5
Cache write, 1 hour $4 $7.20 $8
Opus 5.5 supports a 1M-token context and up to 128K output tokens. Adaptive thinking is always enabled; the default effort is medium. Requests that disable thinking are translated to low effort. Other models and their prices are unchanged. Requests already accepted and historical usage retain their captured prices.
Manual thinking budgets are converted to effort levels. Opus 5.5 does not support forced tool selection (any/tool, or OpenAI required/named functions); these requests return an error without a charge. Use automatic tool selection for supported tool conversations. Provider-native computer-use clients must adopt Anthropic's new toolset format; ordinary custom tools are unchanged.
GPT-6 Sol is available at half the OpenAI Standard price
· improved
GPT-6 Sol replaces GPT-5.6 Sol, with shared routes at 50% and official routes at 90% of Standard list rates. Existing model IDs and keys remain compatible.
Use gpt-6-sol on the shared or official OpenAI route. Calls to gpt-5.6-sol, including reasoning suffixes, now use GPT-6 Sol at the new prices. Keys restricted to the previous model also allow GPT-6 Sol.
USD per million tokens, based on OpenAI's Standard prices, checked September 23, 2026:
Category Shared (50%) Official (90%) OpenAI Standard
Input $1 $1.80 $2
Output $5 $9 $10
Cache read $0.10 $0.18 $0.20
Cache write $1.25 $2.25 $2.50
These rates apply up to 272K input tokens. Above that, the whole request bills at the long-context rates, as on OpenAI: input and cache rates double and output rises by half (input, output, cache read and cache write are $2, $7.50, $0.20 and $2.50 on shared, $3.60, $13.50, $0.36 and $4.50 on official). An earlier version of this note said the prices stayed fixed at every context length; that was wrong, see Long-context prices for GPT-6 Sol and Grok 4.7. Cache writes and reads are separate token categories, not additional fees on top of uncached input. Other models retain their current prices, and accepted requests and historical usage retain their captured prices.
GPT-6 Sol supports text and image input, a 1.05M-token context window, up to 922K input tokens and 128K output tokens. Reasoning levels are none, low, medium (default), high, xhigh and max; legacy minimal maps to low.
Use the Responses API for reasoning and tools. TopxAI also accepts Chat Completions and Anthropic-compatible clients: reasoning-enabled Chat requests use the existing Responses conversion, preserving the client response format and streaming. Chat requests with reasoning_effort: none can use native Chat tool calling. Sampling parameters unsupported during reasoning are removed before forwarding.
The current upstreams support implicit prompt caching. Explicit prompt_cache_breakpoint requests are currently rejected by both suppliers; leave cache selection implicit. The cache-write rate applies only to write tokens actually reported by the upstream. Cache misses are not treated as cache writes.
Grok 4.7 is available at half the xAI Standard price
· improved
Grok 4.7 replaces Grok 4.6 at 50% and 90% of xAI prices, with discounted cache reads and compatibility for existing model IDs and keys.
Use grok-4.7 on either Grok route. Calls to grok-4.6, including explicit reasoning modifiers, now select Grok 4.7. Existing keys restricted to Grok 4.6 continue to work with the upgraded model.
USD per million tokens, compared with xAI Standard prices below 200K input tokens, verified September 23, 2026:
Category Shared (50%) Official (90%) xAI Standard
Input $1 $1.80 $2
Output $3 $5.40 $6
Cache read $0.25 $0.45 $0.50
These rates apply to prompts under 200K tokens. From 200K, the whole request bills at the long-context rates, twice every rate above, as on xAI: input, output and cache read are $2, $6 and $0.50 on shared, $3.60, $10.80 and $0.90 on official. An earlier version of this note said the prices stayed fixed across the context window; that was wrong, see Long-context prices for GPT-6 Sol and Grok 4.7. Cached input is charged at the cache-read rate instead of the regular input rate. Cache misses use the regular input rate; there is no separate cache-write premium. Reasoning tokens use the output rate. Existing administrator overrides and historical charges are preserved.
Grok 4.7 supports text and image input, a 500K-token context window, function calling and structured outputs. Reasoning is always enabled: choose low, medium, high (default) or xhigh.
Chat Completions and Responses both support streaming and tool use. When continuing a Responses conversation, pass all returned output items into the next input, including reasoning items and their encrypted_content. The configured upstreams currently return grok-4.7-build as their response model identifier for requests to grok-4.7; the offered model and captured TopxAI billing model remain grok-4.7.
Long-context prices for GPT-6 Sol and Grok 4.7
· fixed
GPT-6 Sol above 272K input tokens and Grok 4.7 from 200K bill the whole request at long-context rates, as on OpenAI and xAI.
The GPT-6 Sol and Grok 4.7 upgrades on September 23 dropped the long-context tier that GPT-5.6 Sol and Grok 4.6 carried, so long prompts were billed at the standard rates. The tier is back. Once the input context of a request (prompt plus cached tokens) crosses the threshold, every token in that request bills at the long-context rate, as on OpenAI and xAI. Requests below the threshold cost the same as before, and requests accepted before the change settle at the price they were quoted.
GPT-6 Sol, input above 272K tokens (from 272,001), USD per million tokens:
Category Shared (50%) Official (90%) OpenAI Standard
Input $2 $3.60 $4
Output $7.50 $13.50 $15
Cache read $0.20 $0.36 $0.40
Cache write $2.50 $4.50 $5
Grok 4.7, prompt of 200K tokens or more, USD per million tokens:
Category Shared (50%) Official (90%) xAI Standard
Input $2 $3.60 $4
Output $6 $10.80 $12
Cache read $0.50 $0.90 $1
Rates follow the GPT-6 Sol and Grok 4.7 model pages, checked September 23, 2026. The price table on each model page lists both tiers, and the usage log shows which tier a request settled at.
Codex CLI downloads and a one-command installer
· new
A Downloads page lists the official Codex CLI packages through an npm mirror; one command installs Codex and connects it to TopxAI on macOS, Linux or Windows.
Update — September 30, 2026: The downloads page now provides the TopxAI desktop app. For current Codex CLI downloads, installation scripts and manual configuration, see the Codex CLI guide. The announcement below is retained as a historical record.
Codex CLI users who cannot reach GitHub or npm from their network no longer have to.
Downloads
The new Downloads page, linked from the top bar, lists the Codex CLI packages for macOS (Apple silicon and Intel), Windows (x64 and ARM64) and Linux (x64 and ARM64). Every file is the official npm package for that system, served unchanged by the mirror run by Alibaba Cloud (registry.npmmirror.com), and shown with the integrity value npm records for it; the row for your own computer is flagged. The current release is open, older ones are folded underneath.
One command
The page also carries the installer. On macOS and Linux:
curl -fsSL https://ai.topxea.com/install/codex.sh | sh -s -- --key sk-...
On Windows:
$env:TOPXAI_API_KEY = "sk-..."; irm https://ai.topxea.com/install/codex.ps1 | iex
It downloads the current package for your system through the mirror, verifies its integrity, puts codex on your PATH, adds TopxAI as the model provider in ~/.codex/config.toml with gpt-5.6-sol as the default model, stores the key and checks that the key can list models. Signed in, the page fills the command with one of your own keys. Running it again updates Codex and keeps the configuration.
The guide
The Codex CLI page in the docs now covers the whole path: where to get Codex, what the installer changes, the same setup by hand, models, updating and the usual failures.
DeepSeek V4.1 Flash and V4 Pro join the catalogue, with off-peak and peak hours
· new
deepseek-flash and deepseek-v4-pro on one official line at 80% of list in both provider tiers; every usage-log row names the tier it was billed at.
Update, 2026-10-25: deepseek-v4-pro is retired. New requests for that ID return HTTP 403 permission_denied; the launch details below are historical.
October 24, 2026 update: deepseek-flash has been replaced by the private deployment deepseek-flash-Abliterated, with a new ID, fixed prices and a 262,144-token output limit. The original launch details below describe the retired model; DeepSeek V4 Pro retains its official peak/off-peak pricing.
A seventh line joins the catalogue.
Two DeepSeek models
deepseek-flash (DeepSeek-V4.1-Flash, 1M-token context, up to 384K output tokens, text and images in) and deepseek-v4-pro (DeepSeek-V4-Pro-0813) run on the deepseek-official line through Chat Completions. Both are priced at 80% of the DeepSeek list price.
Off-peak and peak hours
DeepSeek publishes two tariffs. Peak hours are 01:00-04:00 and 06:00-10:00 UTC, Monday to Friday, except Chinese public holidays; every other hour, weekends and public holidays are off-peak at half the peak rate. TopxAI applies the same calendar:
Model Off-peak input / cache read / output Peak input / cache read / output
deepseek-flash $0.12 / $0.0024 / $0.48 $0.24 / $0.0048 / $0.96
deepseek-v4-pro $0.528 / $0.0176 / $1.584 $1.056 / $0.0352 / $3.168
The tier is decided when the relay accepts a request and frozen with its price, so a long stream that crosses the boundary keeps the tier it was quoted. Every usage-log row for these models carries an "Off-peak" or "Peak hours" chip, the billing details name the lanes that were charged, the models table shows both tariffs, and the Studio toolbar says which tier is in force before you send. The hosted MCP list_models tool reports the current tier and both tariffs.
Auto-route keys, the blog, three more languages and the help messenger
· new, improved
One auto-route key for every model, a blog, Japanese, Russian and Spanish interfaces, the shared help messenger and the acceptable use policy.
One key can now cover the whole catalogue. When you create a key, pick Auto route: each request goes to the lowest-priced route that serves its model, shared pool before official line, and a single-line model (GLM, Kimi, Jev) lands on its line. The usage log records the route actually used. Per-route keys remain for pinning a line.
Also new:
The keys page lists the base URL https://ai.topxea.com/v1 and every endpoint it serves.
A blog at /blog; the first article is the Jev setup guide. A model covered by an article gets an Introduction link on the price table.
The interface is available in Japanese, Russian and Spanish, next to English and Chinese.
The Support link in the footer and the help button at the bottom right open a messenger shared with topxea.com, so a conversation continues on either site. Email still works: support@topxea.com.
An Acceptable Use Policy at /acceptable-use-policy in all five languages.
Card payments on topxea.com and gift cards
· new, improved
Card top-ups on topxea.com, and gift cards that can be bought on either site and redeemed for TopxAI credit.
Card / local payment in the wallet now settles through topxea.com. Under Add credit, choose the card option, enter an amount between $10 and $10,000 and press Continue to secure checkout. You land on the payment page of topxea.com (Waffo); once the main site confirms the payment, your balance is credited. If the confirmation is slow, Check payment on the order asks for the latest status. Crypto top-ups with USDT or USDC are unchanged.
The wallet also sells gift cards. Under Gift cards, Buy a gift card issues a one-time redemption code for the amount you paid instead of adding credit to your own balance. The code appears in the order details and under My gift cards; any account can redeem it once under Redeem a code. Credit codes can also be listed for sale on topxea.com.
Usage log rows for images and video now itemize the per-image or per-second charge.
GLM-5.3-Abliterated, Kimi K3 and Jev join the catalogue
· new, fixed
GLM-5.3-Abliterated on a private deployment, Kimi K3 at 80% of list and TypeSafe Jev through /v1/systemone with a one-command setup.
October 24, 2026 update: kimi-k3 has been replaced by Kimi K3.1 Pro. Its default TopxAI token prices are unchanged and now equal 120% of the new model’s official reference. Update the model ID and any restricted key allowlist; the old ID returns HTTP 403. The original release record below is preserved.
Three lines join Claude, OpenAI and Grok.
GLM-5.3-Abliterated
A fine-tuned GLM 5.3 with its refusal alignment removed, served from a private deployment rather than the vendor API. Text only, priced above the Z.ai list price: $4 per 1M input tokens, $7 per 1M output. Fixed: an image, audio or file part sent to this model is replaced with a short note and the text is forwarded, instead of the request failing with a 400.
Kimi K3
kimi-k3 by Moonshot AI runs on one official line at 80% of list price: $2.40 input and $12 output per 1M tokens, with a 1M-token context.
Jev
jev-1.13.0 (aliases jev-latest and jev-preview) is TypeSafe's System One model. It answers typed questions about a state with probabilities instead of generating text, through POST /v1/systemone. $0.04494 per 1M input tokens, 107% of TypeSafe's list price; output is free. A key on the TypeSafe route shows a one-command setup under the key:
curl -fsSL https://ai.topxea.com/install/typesafe.sh | sh -s -- --key sk-xxxx
Image and video generation, long-context tiers
· new, improved
GPT Image 2.5 per image, Grok Imagine Video 1.5 per second, and whole-request long-context tiers for OpenAI and Grok models.
Images
gpt-image-2.5-sunburst and gpt-image-2.5-flare replace gpt-image-2 on the image route, through /v1/images/generations and /v1/images/edits. Billing is per picture by size: $0.03 up to 1024x1024 (1K) and $0.05 up to 2048x2048 (2K). An omitted or auto size is pinned to 1024x1024 before the request is priced.
Video
grok-imagine-video-1.5 on the Grok official line: POST /v1/videos/generations, 1 to 15 seconds (default 8), billed per second by resolution: $0.05 at 480p, $0.10 at 720p, $0.20 at 1080p. Poll GET /v1/videos/{request_id}. A failed or expired job is refunded in full, a shorter clip pro rata.
Long context
gpt-6-astra, gpt-5.6-sol and grok-4.6 now carry the providers' long-context tier: above 272K input tokens for OpenAI, from 200K for xAI, the whole request settles at the higher lanes. Claude, GLM and Kimi have no tier. The price table and the usage log show the tier and the per-unit charges.
TopxAI is live: Claude, OpenAI and Grok through one endpoint
· new
The first release: Claude, OpenAI and Grok models on shared pool and official line routes through one OpenAI- and Anthropic-compatible endpoint.
TopxAI is live at https://ai.topxea.com. It relays Claude, OpenAI and Grok models through one OpenAI- and Anthropic-compatible endpoint. Set the base URL to https://ai.topxea.com/v1 and use your TopxAI key.
Models: claude-fable-5-1, claude-fable-5, claude-sonnet-5-5, claude-opus-5, gpt-6-astra, gpt-5.6-sol, gpt-image-2 and grok-4.6.
Two routes per vendor: the shared pool at 50% of list price and the official line at 90%. Both reach the vendor's own models and differ only in routing and price. You pick the route when you create a key; gpt-image-2 has its own image route.
Fixed USD prices per million tokens, charged from a prepaid balance. Every request writes a usage-log row with model, route and charge.
Zero retention: prompts, completions and files pass through in memory and are never written to disk. Logs hold token counts, model names, routes and identifiers.
Upstream errors are masked into a fixed set of public error codes, each with a request id.
Public pages are crawlable, with sitemap and robots.txt.
Models & pricing · Docs · Blog
Other languages: 简体中文