API pricingAPI pricing for GPT-6.2 Astra is reported as 20% cheaper than its predecessor; listed prices: $8 per million input tokens, $0.40 per million cached input tokens, and $40 per million output tokens.
OPENAI · MODEL RELEASE TRACKER
GPT-6.2 Astra
GPT-6.2 Astra is a recently reported OpenAI model (paired with Sol) that, according to an Artificial Analysis evaluation reported by The Decoder, produces much shorter, less-structured answers and admits uncertainty far less often than its predecessors.CURRENT SNAPSHOT4/5 DIMENSIONS WITH DATA
The dimensions that change the decision.
Not established from the available sources.
ModalityLanguage model.
Benchmark performance and rankingReported to rank second in the Artificial Analysis Intelligence Index and to dominate the Pareto frontier for cost-performance; on Agents' Last Exam it scored 5.2 points above its predecessor (GPT-6 Astra) at under a third of the cost and led Anthropic's Claude Fable 5.2 by 5.5 points at ~5% of the cost; on the legal benchmark Harvey LAB-AA it reportedly more than doubled Claude Fable 5.2's pass rate (excerpt truncated in source).
Cross-generation benchmark performanceIn tests with open RISC‑V processor cores, GPT‑6.2 Astra "delivered the best results across model generations," according to the article.
Adaptive black-box red teaming — attack success rate (internal automated tests)OpenAI reported that in internal automated tests, attacks pulling protected reasoning from prior requests fell from 54.0% on GPT-6 Astra to 0.2% on GPT-6.2 Astra, and attacks on active requests fell from 22.9% to 0.9%.
Organization IndexOrganization Index dropped from 70 to 11 in the evaluation, a composite of traits including structure and clarity.
Comparison with Anthropic modelsIn the Artificial Analysis evaluation, Astra (and Sol) scored lower on structure and transparency than Anthropic's Claude Fable 5.2 and Opus 5.5.
Published prompt templates and documentationOpenAI published copy-ready prompt templates and documentation for GPT-6.2 Astra covering writing style, initiative, and coding to address observed behaviors.
RELEASE TIMELINE
Published, source-backed release events only.
VERIFIABLE FACTS
Every value stays attached to a source and date.
RELEASE · Release / announcementDEVELOPER CLAIM
Announced/released on 2026-11-17.
MODALITIES · ModalityDEVELOPER CLAIM
Language model.
CAPABILITIES · Capabilities (professional, coding, presentation fidelity)DEVELOPER CLAIM
Designed for demanding specialized professional work (science, law, finance); trained to handle professional conventions in fields like law, finance, and accounting more reliably; improved coding and computer-use capability; presentation outputs that pull layouts, colors, typography from templates and match exact font sizes.
AVAILABILITY · Published prompt templates and documentationDEVELOPER CLAIM
OpenAI published copy-ready prompt templates and documentation for GPT-6.2 Astra covering writing style, initiative, and coding to address observed behaviors.
PRICING · API pricingDEVELOPER CLAIM
API pricing for GPT-6.2 Astra is reported as 20% cheaper than its predecessor; listed prices: $8 per million input tokens, $0.40 per million cached input tokens, and $40 per million output tokens.
BENCHMARKS · Benchmark performance and rankingDEVELOPER CLAIM
Reported to rank second in the Artificial Analysis Intelligence Index and to dominate the Pareto frontier for cost-performance; on Agents' Last Exam it scored 5.2 points above its predecessor (GPT-6 Astra) at under a third of the cost and led Anthropic's Claude Fable 5.2 by 5.5 points at ~5% of the cost; on the legal benchmark Harvey LAB-AA it reportedly more than doubled Claude Fable 5.2's pass rate (excerpt truncated in source).
CAPABILITIES · Used for chip design/optimization testingDEVELOPER CLAIM
Tested in AI-assisted chip-design/optimization workflows involving open RISC‑V processor cores.
BENCHMARKS · Cross-generation benchmark performanceDEVELOPER CLAIM
In tests with open RISC‑V processor cores, GPT‑6.2 Astra "delivered the best results across model generations," according to the article.
BENCHMARKS · Adaptive black-box red teaming — attack success rate (internal automated tests)DEVELOPER CLAIM
OpenAI reported that in internal automated tests, attacks pulling protected reasoning from prior requests fell from 54.0% on GPT-6 Astra to 0.2% on GPT-6.2 Astra, and attacks on active requests fell from 22.9% to 0.9%.
SAFETY · Tightened safeguards against extracting internal reasoningDEVELOPER CLAIM
OpenAI tightened safeguards for GPT-6.2 Astra to make it much harder to extract protected internal reasoning. In internal automated tests, attacks pulling protected reasoning from prior requests fell from 54.0% (on GPT-6 Astra) to 0.2% on GPT-6.2 Astra; attacks on active requests fell from 22.9% to 0.9%.
CAPABILITIES · More brief and direct by defaultDEVELOPER CLAIM
OpenAI says GPT-6.2 Astra answers more briefly and directly by default than earlier models, and recommends developers review system prompts when moving to the model.
CAPABILITIES · Positioned for the most demanding tasksDEVELOPER CLAIM
OpenAI positions Astra as built for the most demanding tasks.
LIMITATIONS · May carry shorthand from reasoning into final responsesDEVELOPER CLAIM
On long tasks, GPT-6.2 Astra sometimes carries shorthand from its reasoning or other agents' messages into the final response; OpenAI recommends dedicated prompts so final answers are fully self-contained.
LIMITATIONS · Editing conversation history can discard earlier reasoning stepsDEVELOPER CLAIM
In the API, editing conversation history discards earlier reasoning steps, which can affect token costs and response quality.
LIMITATIONS · Sampling parameters removed (temperature, top_p)DEVELOPER CLAIM
Parameters such as "temperature" and "top_p" are removed in the GPT-6.2 Astra API.
RELEASE · Release statusDEVELOPER CLAIM
Described as a new OpenAI model in an Artificial Analysis evaluation reported by The Decoder.
CAPABILITIES · Average answer lengthDEVELOPER CLAIM
Average answer length dropped from 352 to 210 words (a 40% decrease) compared to its predecessor.
CAPABILITIES · Headings per answerDEVELOPER CLAIM
Average number of headings per answer fell from 1.99 to 0.11.
CAPABILITIES · Answer First trait (score)DEVELOPER CLAIM
Answer First score increased from 49 to 89, indicating the model more often addresses the question directly.
LIMITATIONS · Admits uncertainty (score)DEVELOPER CLAIM
The 'Admits Uncertainty' trait score fell from 96 to 24, indicating the model admits uncertainty far less often than before.
BENCHMARKS · Organization IndexDEVELOPER CLAIM
Organization Index dropped from 70 to 11 in the evaluation, a composite of traits including structure and clarity.
BENCHMARKS · Comparison with Anthropic modelsDEVELOPER CLAIM
In the Artificial Analysis evaluation, Astra (and Sol) scored lower on structure and transparency than Anthropic's Claude Fable 5.2 and Opus 5.5.
WHAT CHANGED
Stored passport versions, without reconstructed history.
LIMITATIONS · Admits uncertainty (score)+ The 'Admits Uncertainty' trait score fell from 96 to 24, indicating the model admits uncertainty far less often than before.
CAPABILITIES · Answer First trait (score)+ Answer First score increased from 49 to 89, indicating the model more often addresses the question directly.
CAPABILITIES · Average answer length+ Average answer length dropped from 352 to 210 words (a 40% decrease) compared to its predecessor.
BENCHMARKS · Comparison with Anthropic models+ In the Artificial Analysis evaluation, Astra (and Sol) scored lower on structure and transparency than Anthropic's Claude Fable 5.2 and Opus 5.5.
CAPABILITIES · Headings per answer+ Average number of headings per answer fell from 1.99 to 0.11.
BENCHMARKS · Organization Index+ Organization Index dropped from 70 to 11 in the evaluation, a composite of traits including structure and clarity.
RELEASE · Release status+ Described as a new OpenAI model in an Artificial Analysis evaluation reported by The Decoder.
LIMITATIONS · Editing conversation history can discard earlier reasoning steps+ In the API, editing conversation history discards earlier reasoning steps, which can affect token costs and response quality.
CAPABILITIES · More brief and direct by default+ OpenAI says GPT-6.2 Astra answers more briefly and directly by default than earlier models, and recommends developers review system prompts when moving to the model.
SAFETY · Deployed defenses and mitigations− OpenAI said defenses for GPT-6.2 Astra operate on two levels: per-request measures (targeted training, input and output filters for messages and tool calls, and API blocks preventing older models from reading protected reasoning) and measures against coordinated campaigns (banning thousands to tens of thousands of accounts weekly).
SAFETY · Tightened safeguards against extracting internal reasoning+ OpenAI tightened safeguards for GPT-6.2 Astra to make it much harder to extract protected internal reasoning. In internal automated tests, attacks pulling protected reasoning from prior requests fell from 54.0% (on GPT-6 Astra) to 0.2% on GPT-6.2 Astra; attacks on active requests fell from 22.9% to 0.9%.
AVAILABILITY · Availability (platforms and subscription tiers)− Available via ChatGPT Work, Codex, and the OpenAI API; available to Plus, Pro, Enterprise, and Edu users. Pro-500 subscribers can use Ultrafast mode with the model.
CAPABILITIES · Positioned for the most demanding tasks+ OpenAI positions Astra as built for the most demanding tasks.
AVAILABILITY · Published prompt templates and documentation+ OpenAI published copy-ready prompt templates and documentation for GPT-6.2 Astra covering writing style, initiative, and coding to address observed behaviors.
LIMITATIONS · Sampling parameters removed (temperature, top_p)+ Parameters such as "temperature" and "top_p" are removed in the GPT-6.2 Astra API.
LIMITATIONS · May carry shorthand from reasoning into final responses+ On long tasks, GPT-6.2 Astra sometimes carries shorthand from its reasoning or other agents' messages into the final response; OpenAI recommends dedicated prompts so final answers are fully self-contained.
BENCHMARKS · Adaptive black-box red teaming — attack success rate (internal automated tests)+ OpenAI reported that in internal automated tests, attacks pulling protected reasoning from prior requests fell from 54.0% on GPT-6 Astra to 0.2% on GPT-6.2 Astra, and attacks on active requests fell from 22.9% to 0.9%.
SAFETY · Deployed defenses and mitigations+ OpenAI said defenses for GPT-6.2 Astra operate on two levels: per-request measures (targeted training, input and output filters for messages and tool calls, and API blocks preventing older models from reading protected reasoning) and measures against coordinated campaigns (banning thousands to tens of thousands of accounts weekly).
No public fact values changed.
CAPABILITIES · Used for chip design/optimization testing+ Tested in AI-assisted chip-design/optimization workflows involving open RISC‑V processor cores.
BENCHMARKS · Cross-generation benchmark performance+ In tests with open RISC‑V processor cores, GPT‑6.2 Astra "delivered the best results across model generations," according to the article.