GPT-6.2 Astra API

GPT-6.2 Astra launched on 2026-11-17. OpenAI calls it its most capable model for your hardest projects, from start to finish. It costs $8 input / $40 output per 1M tokens, with cached input at $0.40; GPT-6 Astra was $10 / $50 with $1 cached input. It beats GPT-6 Astra on all nine head-to-head charts in OpenAI’s launch post, yet the first testers’ reviews are sharply split. This page covers the benchmarks and third-party evaluations, how to choose between it, GPT-6.2 Sol and Claude, how to use it in Codex, what to change when migrating, and Wokey pricing.

Price check: official, OpenRouter and Wokey

USD per million tokens. Official and OpenRouter rates were checked by hand on 2026-11-18, excluding tax; the Wokey rate is read from the live price list.

ChannelInputCache readOutput
Official API$8$0.4$40
OpenRouter$8$0.4$40
Wokey$0.548$0.0274$2.74

· Official pricing page · OpenRouter endpoint list

GPT-6.2 Astra at a glance

Release
2026-11-17, together with GPT-6.2 Sol; available the same day in the API, ChatGPT Work and Codex, and listed on OpenRouter (OpenAI, Azure and Amazon Bedrock endpoints)
Positioning
OpenAI’s flagship for complex reasoning, coding and professional work. OpenAI says its code is easier to read, its frontends and 3D designs are better, its tool use is more efficient and its answers are more concise by default. For day-to-day coding OpenAI suggests starting with GPT-6.2 Sol
Context and output
1,050,000-token context (922K max input) and 128K max output; text and image input, text output, no audio or video; knowledge cutoff 2026-05-18
Reasoning effort
The API rejects none. OpenAI suggests starting new integrations at medium and moving none or minimal to low; the launch charts and Codex offer low, medium, high, xhigh and max, and the docs do not yet state a default
Official pricing
$8 input / $40 output per 1M tokens, cached input $0.40, cache writes $10 (30-minute retention) or $20 (12 hours); requests over 272K input tokens are billed at 2× input and cache and 1.5× output for the whole request ($16 / $0.80 / $60); Batch 50% off, Fast 2×, Ultrafast 3× through 2027-02-28 and 6× after
API and tools
Chat Completions, Responses and Batch; tool calling requires Responses. Built-in tools: web_search, file_search, image_generation, code_interpreter, hosted_shell, apply_patch, skills, computer_use, mcp and tool_search. No fine-tuning, Realtime or Assistants, and no dated snapshot
Efficiency
OpenAI says it scores 5.2 points above GPT-6 Astra on Agents’ Last Exam at under a third of the cost per task, and 6.2 points higher on GDPval v2 with 22% fewer output tokens. Artificial Analysis measures $4.00 per task at max effort, against $26.54 for Claude Fable 5.2
Safety rating
The system card rates cybersecurity capability Critical and biological and chemical High, and it is OpenAI’s first external model to meet the High threshold for AI self-improvement. Higher-risk cybersecurity requests get tighter safeguards, which also restrict some legitimate vulnerability research
Model ID and limits
gpt-6.2-astra; top tier 15,000 RPM / 40M TPM; US and EU data residency, with a 10% uplift on regional endpoints

· Sources: OpenAI · OpenAI API · GPT-6.2 Astra · Latest model guide · Codex pricing · System card · Artificial Analysis · Vals AI · VentureBeat

Official benchmarks: GPT-6.2 Astra vs GPT-6 Astra vs GPT-6.2 Sol vs Claude Fable 5.2 vs Opus 5.5

From the nine comparison charts in OpenAI’s launch post. These are vendor-run results, each at the highest effort OpenAI plotted for the model.

BenchmarkGPT-6.2 AstraGPT-6 AstraGPT-6.2 SolClaude Fable 5.2Claude Opus 5.5
Agents’ Last Exam v1.177.6%73.1%73.0%72.8%71.6%
Terminal-Bench-Science 0.177.1%68.1%68.0%73.3%63.3%
GDPval v2 (internal)50.4%44.2%46.9%44.8%43.5%
FrontierCode 2.0 (Cognition)44.2%42.8%43.8%45.7%45.5%
AutomationBench 1.0.652.4%41.4%41.1%51.4%42.5%
BenchCAD 2.1 Core82.5%62.6%73.1%83.5%65.3%
Harvey LAB-AA v1.115.0%8.6%11.1%7.2%4.2%
Vals Science Index v168.4%65.1%57.8%63.7%58.7%
LifeVal86.1%79.8%83.8%——

On Agents’ Last Exam, GPT-6.2 Astra scores 78.3% at xhigh, above its max result; on FrontierCode 2.0, Fable 5.2 scores 49.1% at xhigh. Fable 5.2’s Terminal-Bench-Science score is Artificial Analysis’ mini-swe-agent run (Anthropic reports 74.6% in Claude Code), and its Vals Science Index score is at high effort. OpenAI says its Fable 5.2 Agents’ Last Exam result excludes nine tasks that used a fallback model; cost per task there is $2.89 for GPT-6.2 Astra at xhigh, $8.99 for GPT-6 Astra and $60.05 for Fable 5.2.

Which to use: GPT-6.2 Astra, GPT-6.2 Sol, GPT-6 Astra or Claude

Versus GPT-6 Astra

The price drops from $10 / $50 to $8 / $40, cached input from $1 to $0.40, and the knowledge cutoff moves from 2026-04-30 to 2026-05-18. It leads on all nine of OpenAI’s comparisons, and the Artificial Analysis Intelligence Index rises from 57 to 62. But AA’s Coding Agent Index (Codex harness, max effort) is flat with GPT-6 Astra, and Every’s review found its writing a step back. Use GPT-6.2 Astra for new work, and compare it on your own tasks before migrating.

Versus GPT-6.2 Sol

Sol costs a quarter as much ($2 / $10). On OpenAI’s FrontierCode 2.0 at max effort, Sol scores 43.8% against Astra’s 44.2% at about a fifth of the cost per task; the AA Intelligence Index has Sol at 57 and Astra at 62. OpenAI itself suggests Sol for day-to-day coding and Astra for professional documents, research, 3D and the hardest multistep work.

Versus Claude Fable 5.2 / Opus 5.5

Fable 5.2 has the same $8 / $40 price and still leads the AA Intelligence Index at 64.6 against Astra’s 62.1, but costs $26.54 per task, more than six times Astra; Opus 5.5 ($4 / $20) scores 59.1 at $10.04 per task. Astra leads both on OpenAI’s Agents’ Last Exam and trails them slightly on FrontierCode 2.0. Most early testers still preferred Claude for long coding sessions; pick Astra when cost per task matters or for 3D and document work.

Setting effort

OpenAI suggests starting at medium and going higher only when the quality gain is worth it. AA measures medium at 57.0 on its index, $1.62 per task and about 2.7 seconds to the first answer token; max reaches 62.1 at $4.00 but waits about 148 seconds; xhigh, at 61.6, $2.52 and about 45 seconds, is the middle ground. Reasoning tokens bill as output, so don’t default to max for interactive use.

OpenAI latest model guide

Third-party evaluations and first-day reception

  • Artificial Analysis: It scores 62 on the Intelligence Index at max effort, 5 points above GPT-6 Astra, ahead of Gemini 4 Carbon (60) and Claude Opus 5.5 (59) and behind only Claude Fable 5.2 (65). Running the whole index cost $4,530, or $4.00 per task. Its AA-Omniscience hallucination rate is 59%, so it still tends to guess when it does not know. It ranks first on AA-3D, AA’s interactive 3D benchmark, winning 139 of 152 human head-to-head votes. Its Coding Agent Index is flat with GPT-6 Astra. Every Wokey meter is 6.85% of the official rate, so the same max-effort task costs about $0.27.
  • Vals AI: At max effort it scores 67.66% on the Vals Index, 7th of 41 models, at $10.68 per test. It ranks first on the Vals Science Index (68.44%), Terminal-Bench-Science (71.43%) and Vibe Code Bench (55.16%), and scores 63.13% on Terminal-Bench 4.0.
  • VentureBeat (first-day reception): The headline’s “essentially unusable” is tester Matt Shumer’s line on X; he used Astra as a sub-agent and the supervising Claude models redid its work. Theo Browne had it optimize a Rust compiler: after seven days it had landed no improvement and left the compiler 56% slower, though he praised its code review, 3D, game development and computer use. Ethan Mollick found it hard to steer but able to deliver on an ambitious project. VentureBeat notes these are a few testers on particular assignments on launch day. On X, early tester Pietro Schirano said most of the issues were with the harness and that without it he was flying.

Migrating from GPT-6 Astra and using it in Codex

These changes come from OpenAI’s model page, latest model guide and Codex docs. Most requests only need model set to gpt-6.2-astra, but check the items below first.

  • Reasoning effort cannot be none. Move none or minimal to low, and start new integrations at medium.
  • Remove temperature, top_p and top_logprobs as OpenAI requires; Chat Completions also drops logprobs, and Responses removes message.output_text.logprobs from include.
  • Tool calling requires Responses. Through Wokey, Chat Completions requests are converted to Responses before forwarding, so Chat requests with tools also work.
  • From GPT-6.2 on, the API drops reasoning items when it detects mid-conversation edits to the history (prefix edits), a model downgrade, or a change of end user or organization, which can lower answer quality and add tokens. Keep the prefix unchanged and append new instructions at the end; that also improves cache hits.
  • To raise or lower effort mid-conversation, append a configuration_update item between replies instead of changing the request-level reasoning.effort. Coming from GPT-5.5 or earlier, replace prompt_cache_retention with prompt_cache_options.ttl ("30m", or "12h" in Responses only).
  • OpenAI says it is more concise and direct by default and may propose a next step and wait for confirmation when intent is unclear. The system card shows it is more likely in two tests to omit obstacles that blocked its work, such as a missing file or unavailable search tool, so ask automated runs to report blockers explicitly.
  • Above 272K input tokens OpenAI bills the whole request at 2× input and cache and 1.5× output. Wokey has no such tier: the full 1M context is billed at one rate.
  • In Codex, GPT-6.2 Astra’s default context is 272K; set model_context_window in config.toml to go up to about 1M. To use it through Wokey, add the provider below and put your Wokey API key in the WOKEY_API_KEY environment variable.

Using GPT-6.2 Astra in Codex through Wokey (~/.codex/config.toml)

# ~/.codex/config.toml
model = "gpt-6.2-astra"
model_provider = "wokey"
model_reasoning_effort = "medium"

[model_providers.wokey]
name = "Wokey"
base_url = "https://api.wokey.ai"
env_key = "WOKEY_API_KEY"
wire_api = "responses"
Model ID
gpt-6.2-astra
Vendors
openai
Supply status
Currently callable

Catalog pricing snapshot

Prices are per 1M tokens; request-time pricing controls settlement.

MeterWokeyOfficial referenceSavings
Input$0.548$893%
Output$2.74$4093%
Cache read$0.0274$0.493%
Cache write$0.685$1093%

Official price source: https://developers.openai.com/api/docs/models/gpt-6.2-astra

Model capabilities

Context
1,050,000 tokens
Max output
128,000 tokens
Streaming
Supported
Tools
Supported
Vision
Supported
Supported API forms
/v1/chat/completions · /v1/responses

Send a request

/v1/chat/completions

curl https://api.wokey.ai/v1/chat/completions \
  -H "Authorization: Bearer $WOKEY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"gpt-6.2-astra","messages":[{"role":"user","content":"Hello"}]}'

/v1/responses

curl https://api.wokey.ai/v1/responses \
  -H "Authorization: Bearer $WOKEY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"gpt-6.2-astra","input":"Hello"}'

Use GPT-6.2 Astra in your tools

The gateway is https://api.wokey.ai (OpenAI-compatible clients usually take https://api.wokey.ai/v1), and the model ID is gpt-6.2-astra. Open the guide for the client you use.

Frequently asked questions

What is the difference between GPT-6.2 Astra and GPT-6.2 Sol, and which should I use?

Astra is the flagship and Sol the everyday model. Astra costs $8 / $40 and Sol $2 / $10, four times less. On OpenAI’s FrontierCode 2.0 at max effort, Sol scores 43.8% against Astra’s 44.2% at about a fifth of the cost per task; on Agents’ Last Exam Astra scores 77.6% and Sol 73.0%. The Artificial Analysis Intelligence Index has Astra at 62 and Sol at 57. OpenAI suggests Sol for day-to-day coding and Astra for the hardest professional, research, 3D and long-running work.

How is GPT-6.2 Astra different from GPT-6 Astra?

It is cheaper and scores higher. The price drops from $10 / $50 to $8 / $40 and cached input from $1 to $0.40. It leads on all nine of OpenAI’s comparison charts, scoring 5.2 points higher on Agents’ Last Exam at under a third of the cost per task, and the Artificial Analysis Intelligence Index rises from 57 to 62. AA’s Coding Agent Index is flat, though, and some testers found its writing worse. Neither model accepts none effort, so migrating usually only needs a new model ID.

Is GPT-6.2 Astra better than Claude Fable 5.2 or Opus 5.5?

On the Artificial Analysis Intelligence Index, Fable 5.2 scores 64.6, GPT-6.2 Astra 62.1 and Opus 5.5 59.1, at $26.54, $4.00 and $10.04 per task. Fable 5.2 has the same $8 / $40 price; Opus 5.5 is $4 / $20. On OpenAI’s Agents’ Last Exam, Astra’s 77.6% leads Fable 5.2 (72.8%) and Opus 5.5 (71.6%), while on FrontierCode 2.0 its 44.2% trails both slightly. Most early testers still preferred Claude for long coding sessions; Astra wins on cost per task.

Why do some people call GPT-6.2 Astra “essentially unusable”?

That is tester Matt Shumer’s verdict on X, which VentureBeat used as its headline. The criticism centres on delegated, long-running work: hard to steer, stopping early, terse summaries, and Theo Browne’s seven days of compiler work without a landed improvement. On the other side, Artificial Analysis ranks it first on AA-3D, Theo praised its code review, 3D and game development, and one early tester said on X that most issues were with the harness. OpenAI’s system card admits it more often omits blockers and says Codex now has instructions that reduce this. These are a few launch-day testers, so test it on your own tasks.

How much does GPT-6.2 Astra cost, and how is context over 272K billed?

The official price is $8 input, $0.40 cached input and $40 output per 1M tokens; cache writes cost $10 for 30-minute retention and $20 for 12 hours. When input exceeds 272K tokens the whole request is billed at $16 / $0.80 / $60. Batch is 50% off, Fast mode 2×, and Ultrafast 3× through 2027-02-28 and 6× after. Wokey has no long-context tier: the full 1M context is billed at one rate, $0.548 input / $2.74 output / $0.0274 cached input per 1M tokens.

Why can’t I find GPT-6.2 Astra in ChatGPT? Is it free?

GPT-6.2 Astra is only in ChatGPT Work and Codex, not the regular Chat view; open Work to select it. Codex docs list Plus, Pro, Business, Enterprise and Edu, with Enterprise and Edu off until an admin enables them; Free and Go are not listed. Without a subscription you can call gpt-6.2-astra in the API and pay per token; through Wokey that is $0.548 input / $2.74 output per 1M tokens.

How far do Codex limits go with GPT-6.2 Astra?

OpenAI estimates 6–50 local messages per 5 hours on Plus (GPT-6 Astra: 5–45). Pro currently has no 5-hour limit, though weekly limits may apply. Codex credits drop to 200 input / 10 cached / 1,000 output per 1M tokens (GPT-6 Astra: 250 / 25 / 1,250). Fast mode uses 2× included usage and Ultrafast 6×; among individual plans only Pro $500 gets Ultrafast on Astra, and Enterprise and Edu need an admin to enable it. Some Plus users report losing a large share of their window in minutes, so use GPT-6.2 Sol or a lower effort for routine work. To avoid the 5-hour window entirely, pay per token through the API: through Wokey, $0.548 input / $2.74 output per 1M tokens, charged to your balance, with no 5-hour or weekly limits.

Which reasoning effort should I use with GPT-6.2 Astra? Does it support none?

No, the API rejects none. OpenAI suggests starting at medium and moving none or minimal to low; the launch charts and Codex offer low, medium, high, xhigh and max. Artificial Analysis measures medium at 57.0 on its index with about 2.7 seconds to the first answer token, xhigh at 61.6 and about 45 seconds, and max at 62.1 and about 148 seconds. Use medium for interactive work and consider xhigh for long tasks.

Is there a GPT-6.2 Luna or GPT-6.2 Terra?

Not as of the check date. OpenAI released only GPT-6.2 Astra and GPT-6.2 Sol and has announced no 6.2 Luna or Terra. For a cheaper tier, use GPT-6.1 Terra ($1 / $5) or GPT-6 Luna ($0.10 / $0.50). Luna, Terra, Sol and Astra are tiers from lowest to highest; 6, 6.1 and 6.2 are generations.

How do I use GPT-6.2 Astra in Codex?

ChatGPT Plus, Pro, Business, Enterprise and Edu users can pick it with /model in the Codex app and CLI or run codex --model gpt-6.2-astra, with usage counted against the plan; the default Codex context is 272K, which model_context_window raises. To pay per token, set model to gpt-6.2-astra in ~/.codex/config.toml and add a provider with base_url https://api.wokey.ai, wire_api responses and env_key WOKEY_API_KEY, at $0.548 input / $2.74 output per 1M tokens (official $8 / $40).

How much does the GPT-6.2 Astra API cost, and how much does it save versus the official API?

Wokey lists input at $0.548 per 1M tokens and output at $2.74 per 1M tokens; the official references are $8 and $40. That is 93% lower for input and 93% lower for output. Request-time pricing controls settlement, and the vendor pricing source is linked on this page.

What is the GPT-6.2 Astra API model ID?

The model ID for GPT-6.2 Astra on Wokey is gpt-6.2-astra. Use it exactly as written in the request body's model field or in your client's model setting (for example ANTHROPIC_MODEL in Claude Code, or model in the Codex CLI config.toml); GET https://api.wokey.ai/v1/models also lists it.

Which API forms does GPT-6.2 Astra support, and how do I call it through Wokey?

GPT-6.2 Astra supports /v1/chat/completions (Chat Completions requests) and /v1/responses (Responses requests). Set the API base URL to https://api.wokey.ai, authenticate with a Wokey API key, and use the canonical model ID gpt-6.2-astra; you can start with /v1/responses. The Send a request section includes a curl example for every supported API form.

How do I use GPT-6.2 Astra in Codex CLI?

Codex CLI calls GPT-6.2 Astra over the Responses protocol. Add a model_provider to ~/.codex/config.toml with base_url = "https://api.wokey.ai", wire_api = "responses", and env_key = "OPENAI_API_KEY"; set model to gpt-6.2-astra and put your Wokey API key in OPENAI_API_KEY. The Codex CLI guide has the full config.

What are the context window and maximum output for GPT-6.2 Astra?

The catalog context window is 1,050,000 tokens and the maximum output is 128,000 tokens. These are separate limits, and a request must also satisfy the upstream constraints in effect when it is processed.

How are cache reads and cache writes priced for GPT-6.2 Astra?

Cache read: Wokey lists $0.0274 per 1M tokens and the official reference is $0.4 per 1M tokens. Cache write: Wokey lists $0.685 per 1M tokens and the official reference is $10 per 1M tokens.

How does calling GPT-6.2 Astra through Wokey differ from OpenRouter?

Wokey lists GPT-6.2 Astra at $0.548 input / $2.74 output per 1M tokens, against an official reference of $8 / $40. OpenRouter states it adds no markup on inference and charges its fee when you buy credits (5.5% by card). Wokey carries a curated set of models and has no per-provider routing parameters, so OpenRouter fits better if you need a wider catalog. The OpenRouter alternative page has the full comparison.

How do GPT-6.2 Astra and GPT-6 Astra differ in price and context?

In the published price comparison, GPT-6.2 Astra lists input/output at $0.548 / $2.74 with a 1,050,000-token context window; GPT-6 Astra lists $0.55 / $2.75 with a 1,050,000-token context window. This is a factual price-and-capacity comparison, not a quality ranking; also compare their supported API forms and capabilities.

Related models and documentation

Get API key · gpt-6.2-astra