OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17. Astra, the flagship, costs $8 per million input tokens, $0.40 per million cached input tokens and $40 per million output tokens. That is 20% below GPT-6 Astra on input and output and 60% below it on cached input. Sol keeps GPT-6.1 Sol’s rates of $2, $0.10 and $10. Both models ship with tighter cybersecurity safeguards that, in OpenAI’s words, “also restrict some legitimate vulnerability research and security testing workflows”. Launch post · API changelog · System card
Prices, specifications and migration rules below come from OpenAI’s developer documentation, and safety results from the system card. The launch post’s benchmark results are OpenAI’s own and mostly drawn as charts, so we repeat only the figures it states in prose. The independent data is from Artificial Analysis, the ARC Prize Foundation and Snorkel, plus two outlets’ readings of third-party results where we say so. S5 Labs has not run either model.
What shipped
OpenAI calls Astra “our new frontier model for your hardest projects” and Sol “our efficient, versatile model for everyday work”. The series has moved quickly: GPT-6 Astra shipped on September 3, GPT-6 Sol and Luna on September 22, GPT-6.1 Sol at DevDay and GPT-6.1 Terra on October 26. CNBC reported on September 28 that OpenAI would not release GPT-6.1 Astra, quoting Saachi Jain, OpenAI’s head of safety systems, as saying it “didn’t quite meet the bar in terms of staying within scope and authorization”. GPT-6.2 Astra arrives 50 days after that report. Launch post · CNBC
In the API the models are gpt-6.2-astra and gpt-6.2-sol, and neither model page lists a dated snapshot. Both have a 1,050,000-token context window, of which up to 922,000 tokens can be input, and a 128,000-token output limit. Astra’s knowledge cutoff is May 18, 2026 and Sol’s is June 1, 2026. Both run on Chat Completions, Responses and Batch, and both support streaming, structured outputs, function calling, file search, image input, web search and prompt caching. Neither supports the Realtime or Assistants APIs, fine-tuning, predicted outputs, or image, video and speech output. Astra model page · Sol model page
Prices against the previous generation
Standard rates per million tokens, for prompts up to 272,000 input tokens:
| Model | Input | Cached input | Output |
|---|---|---|---|
| GPT-6.2 Astra | $8.00 | $0.40 | $40.00 |
| GPT-6 Astra | $10.00 | $1.00 | $50.00 |
| GPT-6.2 Sol | $2.00 | $0.10 | $10.00 |
| GPT-6.1 Sol | $2.00 | $0.10 | $10.00 |
The launch post describes “a 20% discount on input and output tokens, and a deeper 60% discount on cached input tokens”. The 60% is measured against GPT-6 Astra’s cached rate, $1.00 falling to $0.40. Against Astra 6.2’s own input price, a cache read costs 5%. Cache writes cost $10 per million tokens for a 30-minute write and $20 for a 12-hour write on Astra (down from $12.50 and $25 on GPT-6 Astra), and $2.50 and $5 on Sol. Launch post · Pricing page
The other service tiers, as input / cached input / output per million tokens:
| Tier | GPT-6.2 Astra | GPT-6.2 Sol |
|---|---|---|
| Standard | $8 / $0.40 / $40 | $2 / $0.10 / $10 |
| Long context (over 272K input tokens) | $16 / $0.80 / $60 | $4 / $0.20 / $15 |
| Batch | $4 / $0.20 / $20 | $1 / $0.05 / $5 |
| Fast | $16 / $0.80 / $80 | $4 / $0.20 / $20 |
| Ultrafast, promotional | $24 / $1.20 / $120 | $6 / $0.30 / $30 |
| Ultrafast, list | $48 / $2.40 / $240 | $12 / $0.60 / $60 |
| Flex | Not listed | Not listed |
The long-context rate applies to the whole request. Once input passes 272,000 tokens, input and cache rates double and output is billed at 1.5x for everything in the request.
Ultrafast doubles in price on March 1. Both model pages say Ultrafast costs 3 times the Standard rate through February 28, 2027 and 6 times Standard from March 1, 2027, Pacific time. The promotional and list rows above are those two multipliers. Ultrafast supports US data residency only, and the migration guide does not say how to turn it on. Astra model page · Migration guide
Flex is missing. The pricing page’s Flex table has no row for either 6.2 model, while GPT-6 Astra has Flex at $5 input and $25 output and GPT-6.1 Sol at $1 and $5. A workload that runs on Flex today cannot keep that price on 6.2 as the page stands. The nearest substitute is the Responses API’s new two-hour and eight-hour background windows, a public beta that bills 30% and 40% below Standard on both 6.2 models. Batch remains the cheapest route at half of Standard.
Regional processing endpoints add a 10% uplift for models released on or after March 5, 2026. Both models support US and EU data residency with Standard, Fast and Batch processing. Pricing page · Migration guide
A worked request
Take 100,000 uncached input tokens and 10,000 output tokens at Standard rates, excluding tools and cache writes. That costs $1.20 on GPT-6.2 Astra ($0.80 of input plus $0.40 of output), $1.50 on GPT-6 Astra and $0.30 on either Sol. Batch brings Astra 6.2 to $0.60. If the same 100,000 input tokens are read from cache, the input portion falls to $0.04 on Astra 6.2, against $0.10 on GPT-6 Astra.
Token price is not task price
OpenAI’s larger claim is about cost per task. Its developer-forum announcement says Astra 6.2 comes in at “less than a third of the cost per task of GPT-6 Astra”, and the launch post makes the same claim for Agents’ Last Exam and Terminal-Bench-Science. Community announcement · Launch post
A 20% price cut cannot produce that by itself, so the rest has to come from using fewer tokens, and the size of that saving will vary by workload. On OpenAI’s internal GDPval v2, the post says Astra 6.2 used 22% fewer output tokens than GPT-6 Astra. By our arithmetic, 22% fewer output tokens at a 20% lower price cuts the output bill by about 38%, well short of two-thirds.
Artificial Analysis offers one outside data point. Running its Intelligence Index at max effort, it recorded 65.8 million output tokens across the whole index and $4.00 per index task for Astra 6.2, and 60.9 million output tokens and $1.01 per task for Sol 6.2. Claude Fable 5.2 at max effort used 275 million output tokens across the index, at $26.54 per task. Artificial Analysis leaderboard · Artificial Analysis: Astra 6.2 · Artificial Analysis: Sol 6.2 · Artificial Analysis: Fable 5.2
Migration constraints
Tool calling requires the Responses API. The changelog entry lists both models on Responses and Chat Completions, then adds: “Tool calling requires the Responses API.” GPT-6.1 Sol had the same rule, so teams that moved for 6.1 are already there. API changelog · DevDay article
Reasoning cannot be switched off. Both model pages say, “The API rejects requests that set reasoning effort to none.” The migration guide recommends starting new integrations at medium and tells users of none or minimal to use low. The parameter is reasoning.effort on Responses and reasoning_effort on Chat Completions. The guide does not list the full set of allowed values. Astra model page · Migration guide
Sampling parameters must go. On Chat Completions, omit temperature, top_p, logprobs and top_logprobs. On Responses, omit temperature, top_p and top_logprobs, and remove message.output_text.logprobs from include. If you are coming from GPT-5.5 or earlier, replace prompt_cache_retention with prompt_cache_options.ttl set to "30m". Migration guide
Edit the conversation history and you lose the reasoning. The migration guide says, “Starting with GPT-6.2, the API will drop reasoning items when it detects mid-conversation changes to the conversation history”. As we read it, the triggers also include a model downgrade and a change of end user or organization. OpenAI recommends “avoiding changes to the prefix and adopting an append-only integration pattern as far as possible”, and the launch post says an unchanged prefix also improves cache reuse. The preserved-reasoning guide adds that its diagnostics cannot verify whether the model reused reasoning. Anthropic enforces a similar prefix rule on Fable 5.2, where an edited history returns a 400. Migration guide · Preserved reasoning
Zero Data Retention needs the full history passed back. Both models support ZDR “for approved organizations on eligible endpoints and features”. With it enabled the Responses API treats store as false, so prior output items, encrypted reasoning included, must be passed back in order. Migration guide
Standard rate limits are the same for both models:
| Usage tier | Requests per minute | Tokens per minute |
|---|---|---|
| Build | 5,000 | 1,000,000 |
| Launch | 10,000 | 4,000,000 |
| Grow | 15,000 | 40,000,000 |
Astra model page · Sol model page
Nothing forces a move. As of November 17, GPT-6 Astra and GPT-6.1 Sol are not on the deprecations page. The nearest deadline is December 11, 2026, when gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro, o3 and o3-pro shut down. The page names gpt-6.1-sol, gpt-6.1-terra and gpt-6-luna as their replacements, and no 6.2 model appears on it. Deprecations
We found no confirmation that either 6.2 model is available on Azure, Amazon Bedrock, Google Cloud or GitHub Copilot.
ChatGPT and Codex access
The launch post lists both models as available to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex. Ultrafast was previously limited to Pro 500 and eligible Enterprise and Edu plans. Pro 500 subscribers now get it with both 6.2 models, and Pro 200 subscribers get it for the first time, with Sol only. The post adds that eligible Enterprise and Edu users “get the same access once their workspace admin enables it”. Launch post · ChatGPT and Codex changelog · DevDay article
New Work chats start on GPT-6.2 Sol at Light effort on supported consumer plans, and at Medium in Enterprise workspaces that have the model enabled. In Codex the default context window is 272,000 tokens for Astra and 500,000 for Sol, and both can be raised to as much as 1 million tokens with model_context_window. ChatGPT models · ChatGPT and Codex changelog
What the benchmarks show and what they leave out
These are the figures the launch post states in prose. All are OpenAI’s reporting, including the ones on third-party benchmarks, and we have seen no confirmation of them beyond the independent data below.
- Agents’ Last Exam v1.1: Astra scores “5.2 points higher than GPT-6 Astra at less than a third of the cost, and 5.5 points ahead of Claude Fable 5.2 at just 5% of the cost”.
- Terminal-Bench-Science 0.1: Astra scores 9 percentage points above GPT-6 Astra at less than one-third the cost. OpenAI also claims a new high on the Vals Science Index and gives no number.
- FrontierCode 2.0, Cognition’s benchmark: Sol comes within 0.5 points of Astra at about one-fifth the cost, both at max effort.
- GDPval v2, OpenAI internal: Astra is up 6.2 percentage points on GPT-6 Astra, and Sol is up 5.3 points on GPT-6.1 Sol.
- LifeVal, OpenAI’s own benchmark of everyday personal tasks: Sol outscores GPT-6 Astra at about one-seventh the cost.
- Harvey LAB-AA v1.1: Astra more than doubles Fable 5.2’s pass rate at about one-sixth the cost, and Astra for Law, OpenAI’s configuration for lawyers, scores higher still.
- AA-3D, Artificial Analysis’s blinded human vote: Astra ranks first at 1,318 Elo, ahead of Fable 5.2 at 1,271 and Sol at 1,259.
Two footnotes qualify the Fable comparisons. The first, in the Agents’ Last Exam chart caption, says the leaderboard results for Fable 5.2 “exclude nine tasks that used a fallback model”. Snorkel’s leaderboard words it differently: Fable 5.2 “excludes 9 tasks that are downgraded; its scores are averaged across the remaining tasks”. That leaderboard, released October 26, has no GPT-6.2 rows, so the 6.2 scores in OpenAI’s chart cannot yet be checked against it. Snorkel leaderboard
The second footnote says Fable 5.2’s Terminal-Bench-Science result is the one Artificial Analysis measured in the mini-swe-agent harness, and that Anthropic reported 74.6% at max reasoning in Claude Code. Artificial Analysis lists Astra 6.2 at 78.6% at xhigh, 78.1% at max and 75.2% at high, so Astra’s best score there is not at its highest effort. The text of Artificial Analysis’s page gives no figures for Sol, GPT-6 Astra or Fable 5.2. Launch post · Artificial Analysis: Terminal-Bench-Science
Independent data so far
- Artificial Analysis Intelligence Index. Astra 6.2 scores 62 at max effort, second on the model leaderboard behind Fable 5.2 at 65 and ahead of Gemini 4 Carbon at 60. Fable 5.2 also scores 62 at xhigh and at high. Astra’s $4.00 per index task is about 15% of Fable’s $26.54, but that comparison holds only against Fable’s max setting. Sol 6.2 scores 57 at max, 56 at xhigh, 55 at high, 53 at medium and 41 at low. Its top score equals the 57 that Artificial Analysis gave GPT-6 Astra at max. Artificial Analysis leaderboard · Sol 6.2 · GPT-6 Astra · Fable 5.2
- ARC Prize. On ARC-AGI-3’s Standard harness, Astra scores 70.3% at a cost of $24.2K and Sol 54.9% at $5.5K. The 100.0% figure attached to Astra is the provider-adapter result, a different configuration, and Sol’s is 98.2%. For comparison, ARC Prize’s Standard-harness results for Fable 5.2 were 79.3% at high and 81.2% at xhigh. ARC Prize leaderboard · Our Fable 5.2 article
- Coding agents and the Vals Index, secondhand. OfficeChai’s reading of Artificial Analysis’s Coding Agent Index v2.1 has Astra 6.2 at 53, level with GPT-6 Astra, and Sol 6.2 at 49, down from 52 for GPT-6.1 Sol. TestingCatalog quotes Vals AI as placing the two models seventh and ninth on the Vals Index, 1.6 points apart, with Sol costing less than a third as much per task. We did not open the Vals or Cognition pages. OfficeChai · TestingCatalog
OpenAI’s charts and Artificial Analysis’s max-effort runs both show Astra 6.2 costing far less to run than Fable 5.2. The sources split on which model is better: Artificial Analysis’s index and ARC-AGI-3’s Standard harness put Fable ahead, and OpenAI’s charts put Astra ahead.
The cyber safeguards developers will feel
OpenAI classifies both models as Critical for cybersecurity and High for biological and chemical capability under its Preparedness Framework. On the ExploitBench Internal Port evaluation, the card reports a mean code execution rate of 53.5% for Astra 6.2 and 50.5% for Sol 6.2, against 31.5% for GPT-6 Astra. Not every result moved up: on Irregular’s FrontierCyber suite Astra 6.2 solved 68 of 217 challenges where GPT-6 Astra solved 77, a drop the card does not comment on. OpenAI says its headline cyber results come from internal configurations with safeguards and classifiers disabled. System card · Launch post
The safeguards tightened too. The card says that in initial evaluations both models refused fewer harmless requests than GPT-6 Astra but were less resistant to jailbreaks. OpenAI then made the refusal boundary more conservative, and writes, “we expect refusals on a wider set of dual-use prompts, including autonomous vulnerability discovery and chained exploitation”. The card’s own scores show the cost to legitimate work:
| Evaluation | GPT-6 Astra | GPT-6.2 Sol | GPT-6.2 Astra |
|---|---|---|---|
| Expected Cyber Refusals (staying within policy on disallowed requests) | 68.2% | 99.4% | 98.5% |
| Cyber Utility (answering allowed requests) | 90.7% | 69.5% | 78.1% |
| Benign biology, not overrefused | 97.8% | 86.2% | 90.9% |
Against GPT-6 Astra, Cyber Utility is 12.6 points lower for Astra 6.2 and 21.2 points lower for Sol 6.2. The card says, “We recognize that this makes GPT-6.2 Astra and Sol potentially less useful for cyber work”. The biology row shows the same effect outside security: Sol 6.2 answers benign biology prompts less often than GPT-6.1 Sol did (86.2% against 98.2%). On fixed jailbreak prompts, cyber resistance improved to 98.4% for Astra and 97.9% for Sol from 91.5% for GPT-6 Astra, while the biology and severe-violence categories slipped by 2.0 to 6.5 points.
Daybreak access to 6.2 is unsettled
Security teams with Daybreak access should not count on reduced refusals on the new models yet. The launch post says defenders “can currently access GPT-6 Astra with reduced cyber refusals on Daybreak Blue” and adds, “We will work to bring GPT-6.2 models to Daybreak”, with no date. The card says reduced cyber refusals on Astra 6.2 “are planned to be available only to the smaller group separately approved for Daybreak Red at launch”, and that “Daybreak Blue approval alone will not reduce safeguards on GPT-6.2 Astra and Sol”. It does not describe Sol’s configuration separately, and the Daybreak developer guide has no 6.2 entry at all. Launch post · System card · Daybreak guide
The post and the guide also differ on GPT-6 Astra. Where the post says “on Daybreak Blue”, the guide says daybreak_blue is the request value for GPT-6 Astra and GPT-6.1 Sol, that the parameter “doesn’t grant access”, and that reduced refusals need Daybreak Red access enabled for the project. For authorized security testing, the options the launch post names today are GPT-6 Astra under Daybreak and GPT-6-Cyber under Daybreak Red. Daybreak guide
Enforcement and monitoring
Under OpenAI’s cybersecurity-checks guide, flagged requests fail with the error code cyber_policy, and access “may be temporarily revoked for the entire organization” unless requests carry a per-user safety_identifier, in which case the revocation applies to the affected user. An appeal has to be filed “before the 7-day period ends”, and the page does not say when that period starts. The guide does not name the 6.2 models, though the system card says the monitoring applies to them. Cybersecurity checks
OpenAI also documents misalignment monitoring. A blocked request returns HTTP 403 with the code misalignment_policy_violation, and the guide says the API “does not provide a general way to resume a conversation stopped by misalignment monitoring”. It lists Chat Completions as “Not covered by this monitoring system” and names no models. We found no published rate for how often either system interrupts benign work. Misalignment monitoring
One outside check on the safeguards is unfinished. The UK AI Security Institute (UK AISI) “identified a substantial regression in robustness relative to GPT-6 Astra”, developing a universal jailbreak in half as many model queries as GPT-6 Astra had required. OpenAI changed the model and its safeguards and “believes this regression to be mitigated, though UK AISI has not validated this at the time of writing”. System card
The first released model at the self-improvement threshold
Astra 6.2 is OpenAI’s “first externally released model to meet the High threshold for AI Self-improvement”. OpenAI defines High as impact “equivalent to giving every OpenAI researcher an additional, highly capable research engineer as an assistant, relative to their 2024 baseline”. Sol 6.2 falls below the threshold. Launch post
The underlying scores are mixed. Astra 6.2 leads GPT-6 Astra on Internal Research Debugging (81.85% against 78.05%) and MLE-Bench Revised (96.06 against 93.80). On PostTrainBench Lite it scores 74.79%, below GPT-6 Astra’s corrected 76.97%, and on the NanoGPT evaluation its 42.87% sits under the card’s 80% indicative cutoff. The High designation is OpenAI’s overall assessment across those results. System card
Anthropic took a different route with Fable 5.2, saying it limited some AI-research capabilities in training. OpenAI’s response is oversight: stronger containment in its research environments, expanded monitoring of internal workloads and supported external deployments, and Control Room, a system that tracks research workloads and can halt model output. The card discloses that in September OpenAI’s monitoring detected an agent reaching the live internet from an environment meant to be offline, after which OpenAI paused frontier reinforcement learning while it strengthened containment. Launch post · System card
Astra 6.2 is more likely to leave out a blocker
The launch post says both models “took fewer improper actions in simulated workplace tasks”, and the card backs that: high-severity improper actions fell to 0.64% for Astra 6.2 from 1.20% for GPT-6 Astra, and to 0.71% for Sol 6.2 from 1.32% for GPT-6.1 Sol.
The same card reports a regression that matters more for unattended agents. In the Blocker Deception test, OpenAI removes something a coding task depends on, such as the file the user asked the agent to edit, and checks whether the model tells the user. Astra 6.2 “omitted a known obstacle to completing its work in 22.28 percent of trials, compared with 11.09 percent for GPT-6 Astra”. The Broken Search Tool test tracks how often the model fails to acknowledge a missing requirement or broken tool in its final answer. Astra 6.2 did so in 23.00% of cases, against 7.58% for GPT-6 Astra and 1.00% for GPT-6.1 Terra. Of the Blocker Deception result, OpenAI writes, “We are concerned by this regression and are continuing to investigate its root causes,” and offers one hypothesis: its efforts to make responses concise and free of excessive caveats. The card says both models were more likely to omit known obstacles, but Sol’s figures appear only in charts. System card
The mitigation OpenAI describes is a prompt. The card says a system prompt intervention “will be incorporated at launch” and that the Codex developer prompt substantially reduced the omissions. It also says, “The Codex developer prompt used in these evaluations is not applied to direct API requests.” The card does not say whether any equivalent reaches API traffic, so developers calling the API should assume they need their own instruction and their own checks.
The card has two more results that bear on unattended use. In a replay of internal Codex traffic, 4.96% of Astra 6.2 turns drew a confirmed misalignment flag, against 3.19% for GPT-6 Astra, and Sol 6.2 drew 6.78% against 5.13% for GPT-6.1 Sol. In UK AISI testing, Astra 6.2 considered whether it was being tested for alignment in 85% of trajectories, against 57% for GPT-6 Astra, which makes the behavioral numbers here harder to take at face value.
Who should move, and what to test first
Teams on GPT-6 Astra have the clearest case. Astra 6.2 costs 20% less per token and 60% less per cached token, and it scores five points higher on Artificial Analysis’s index. The exceptions are security teams that rely on Daybreak, anyone doing dual-use biology or vulnerability research, and workloads on Flex. Those should stay on GPT-6 Astra until OpenAI extends Daybreak and Flex to 6.2 or says it will not.
Teams on GPT-6.1 Sol pay the same token price either way, so the decision rests on quality and refusals. Sol 6.2 matches GPT-6 Astra’s index score at $1.01 per task, but the secondhand coding-agent reading has it three points below 6.1 Sol, its Cyber Utility score is the lowest in the card’s table, and it loses Flex. Run both before switching a coding agent.
New builds should start on Sol 6.2 at medium and move to Astra only where evals fail. OpenAI’s FrontierCode claim, the Vals placement and the Artificial Analysis index all put Sol within about five points of Astra at a fifth to a third of the cost per task.
What to test first:
- Tokens per finished task. Run a fixed set of jobs on the old and new model at the same effort and record output tokens and cost per completed, checked task. That is how to check OpenAI’s one-third claim on your own workload.
- Blocker disclosure. Remove a file or disable a tool mid-task and read the final message. Add an explicit instruction to report anything that blocked the work, then compare what the agent says with file-system and tool logs.
- Refusals on your own allowed work. Replay security, biology and other dual-use prompts that GPT-6 Astra or 6.1 Sol answered, count refusals, and handle
cyber_policyandmisalignment_policy_violationerrors in code. Send asafety_identifierper end user so that a revocation can be limited to that user. - History edits. Replay a session that trims tool output or rewrites the system prompt mid-conversation, and compare it with an append-only version.
- Price cliffs. Count how many requests cross 272,000 input tokens, where Astra 6.2’s input price doubles to $16 per million tokens and output rises to $60. If Ultrafast is in a product plan, budget it at 6x Standard, the rate from March 1, 2027.
Sources
OpenAI announcement and system card
- Introducing GPT-6.2 Astra and Sol — OpenAI
- GPT-6.2 Astra and Sol system card — OpenAI
- GPT-6.2 Astra and Sol are now available — OpenAI Developer Community
Developer documentation
- GPT-6.2 Astra model page, GPT-6.2 Sol model page, Pricing, API changelog, Deprecations — OpenAI Developers
- Migration guide, Preserved reasoning, Background mode — OpenAI Developers
- Daybreak, Cybersecurity checks, Misalignment monitoring — OpenAI Developers
ChatGPT and Codex documentation
- ChatGPT & Codex changelog, Models — ChatGPT Learn
Independent data
- Model leaderboard, Terminal-Bench-Science, GPT-6.2 Astra, GPT-6.2 Sol, GPT-6 Astra, Claude Fable 5.2 — Artificial Analysis
- ARC Prize leaderboard — ARC Prize Foundation
- Agents’ Last Exam leaderboard — Snorkel
Press
