
GPT-6.2 Astra and Sol: Same Launch, Opposite Verdicts on Coding
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 170 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 234 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 89 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 98 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-079 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 88 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 134 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 79 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 94 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 118 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 68 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 94 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 44 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 199 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 114 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 311 tok/s
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
Two models went live on 17 November and the cheaper one arrives with a regression attached. GPT-6.2 Sol scores 57 on the Artificial Analysis Intelligence Index at $1.01 per index task; GPT-6.2 Astra scores 62 at $4.00. Sol is the tier OpenAI tells you to make your default, the one its own staff reportedly reach for on ordinary work. And on Artificial Analysis's Coding Agent Index — the same Codex harness, at max effort — GPT-6.2 Sol lands three points below the model it replaces, GPT-6.1 Sol, with the drop traced to a single evaluation where it omits key facts 14 points more often. Astra, measured the same way, is flat against GPT-6 Astra. One launch, two opposite verdicts on the thing Codex users care about most.
That split is worth an article because the announcement does not make it. OpenAI's post is about slides, spreadsheets, 3D scenes and scientific terminals, and on those it has real numbers. The coding picture is the one you have to assemble from independent measurements, a changelog entry and a migration guide, and it changes which of the two IDs belongs in your repository.
What shipped, precisely
The changelog entry for 17 November lists two model IDs, gpt-6.2-astra and gpt-6.2-sol, on both the Responses API and Chat Completions. Tool calling requires Responses. In ChatGPT, both arrive for Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex on day one; Pro 500 subscribers get Ultrafast on either model, and Ultrafast reaches Pro 200 for the first time, starting with Sol. Eligible Enterprise and Edu workspaces get access once an admin enables it.
• Price — GPT-6.2 Astra $8 input / $40 output per million vs GPT-6.2 Sol $2 / $10
• Cached input — $0.40 per million on Astra vs $0.10 on Sol
• Cache writes — $10 per million for a 30-minute retention window on Astra vs $2.50 on Sol
• Context window — 1,050,000 tokens for both, of which 922,000 max input and 128,000 max output
• Knowledge cutoff — 18 May 2026 on Astra vs 1 June 2026 on Sol, so the cheaper model is the more current one
• Modalities — text and image in, text out, on both; audio and video unsupported on both
• Tools — web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, MCP and tool search, all on the Responses API

The price move is asymmetric and easy to misread. Astra sits 20% below GPT-6 Astra on both input and output, and cached input falls from $1.00 to $0.40 — a 60% cut, because the discount deepened to 95% and the base rate it applies to fell at the same time. Sol did not move at all: $2, $0.10 and $10 are exactly GPT-6.1 Sol's rates. So the cheaper tier is a capability upgrade at a constant price, and the expensive tier is a capability upgrade with a price cut stapled to it. Which of those is the better buy depends entirely on what you are buying it for.
Astra is the ceiling, and for once the ceiling is measurable
Artificial Analysis places GPT-6.2 Astra third of 91 models on its Intelligence Index at 62, behind Claude Fable 5.2 at 65. It clears Gemini 4 Carbon at 60 and Claude Opus 5.5 at 59, which is a real move: the previous GPT-6 Astra sat five points lower. The index is an average of nine evaluations covering knowledge work, agentic terminal use, science and general knowledge, so a five-point gain on it is not one benchmark being gamed.

ARC Prize's independent verification is the most striking single number in the launch. GPT-6.2 Astra is the first model to score 100% on ARC-AGI-3, at a total cost of $4.1K through a provider adapter; under ARC's standard configuration the same model scores 70.3% at $24.2K. It also reports 95.8% on ARC-AGI-2 at $0.50 per task and 98.5% on ARC-AGI-1 at $0.17 per task. Those are third-party measurements on a benchmark family built to resist memorisation, not vendor claims, and they are the strongest evidence in this release that the model does something its predecessor did not. The cost spread across the two ARC-AGI-3 configurations is the caveat: a perfect score is available, and how much of it you get depends on how much scaffolding you are willing to pay for.
OpenAI's own figures, which are vendor-reported and unreproduced, point the same direction: 5.2 points above GPT-6 Astra on Agents' Last Exam at under a third of the cost, 9 percentage points above it on Terminal-Bench-Science 0.1, a new high on the Vals Science Index, and a state of the art on Harvey LAB-AA v1.1 for legal work. Artificial Analysis independently has Astra leading AA-3D at 1,318 Elo against Claude Fable 5.2 at 1,271, and leading AA-Robotics v0.1 at 79. It is second only to Fable 5.2 on factual knowledge accuracy and on CritPt v2's physics subset.
The counterweight is cost and latency, and both are worse than the index suggests. Its $4.00 per index task ranks 61st of 90 on Artificial Analysis's cost ordering, nearer the expensive end than a third-place intelligence rank prepares you for, and it takes 10.4 minutes per task with a time to first answer token of 134.86 seconds. That last figure is what an interactive Codex session feels like: two minutes of nothing before the first word. Astra is a model you point at a hard problem and walk away from.
Sol is where the volume goes, and where the regression lives
Sol's case is arithmetic. 57 on the same index puts it eleventh of 91 rather than third, level with the previous flagship GPT-6 Astra, at roughly 15% of that model's measured cost per task. Artificial Analysis puts Sol at $1.01 per index task against Astra's $4.00 — a factor of four — while finishing a task in 7.9 minutes rather than 10.4 and reaching a first answer token in 57.55 seconds rather than 135. It is faster, cheaper and less verbose, spending 60.9 million output tokens across the index suite against Astra's 65.8 million.
OpenAI's framing is that Sol is the everyday default, and on several of its internal evaluations that holds. On Cognition's FrontierCode 2.0, Sol comes within half a point of Astra at about a fifth of the cost, both at max effort. On LifeVal, a new evaluation for practical personal tasks, it outscores GPT-6 Astra at roughly a seventh of the price. On the internal GDPval v2 knowledge-work suite it improves 5.3 points over GPT-6.1 Sol.
Then there is the coding measurement that runs the other way. On Artificial Analysis's Coding Agent Index, run in the Codex harness at max effort, GPT-6.2 Astra is flat with GPT-6 Astra — the upgrade buys nothing measurable — and GPT-6.2 Sol is three points below GPT-6.1 Sol. Artificial Analysis attributes the Sol drop to a 14-point fall on SWE-Atlas-QnA-AA, an evaluation where the model has to answer questions about a codebase, and describes the cause as the model being more likely to omit key facts in its responses.
That is a specific, checkable failure mode, and it is the one that hurts an agent. An agent that leaves out a fact it found does not crash; it produces a plausible edit that misses the reason the code was written that way. If your Codex usage is "ask questions about this repository and act on the answers," the cheaper model in this pair is not an upgrade on the dimension you are buying. If it is "write this function, run the tests, fix what breaks," the FrontierCode result says you will not notice the difference and will notice the invoice.
The wiring, which is where a migration actually breaks
Three mechanics matter more than the benchmarks once you move a real integration.
The first is that both models reject reasoning_effort: "none". GPT-6 Sol accepted it; GPT-6.1 Sol did not, and GPT-6.2 continues that. If you were relying on turning reasoning off to get latency back on a cheap classification path, that path does not exist on this pair. You buy latency with a smaller model or a shorter prompt, not with a flag.
The second is the long-context cliff. Above 272,000 input tokens, input and cached-input rates double and output costs 1.5 times the short-context rate — and the multiplier applies to the whole request, not just the portion past the threshold. On Astra that turns $8 and $40 into $16 and $60 the moment a session crosses the line; on Sol, $2 and $10 into $4 and $15. A 1,050,000-token window invites exactly the behaviour that triggers it.
The third is that the price card has a time bomb in the service tiers. Batch bills at 50% of standard. Fast mode bills at double. Ultrafast bills at three times standard until 28 February 2027 and six times from 1 March 2027, Pacific time. Anyone who adopts Ultrafast at launch and hard-codes today's cost model into a budget should know the rate doubles in about three months. There is also a discount worth taking: background completion windows in the Responses API, still in public beta, bill model token rates 30% below standard for the two-hour window and 40% below for the eight-hour window, and both GPT-6.2 models are eligible.
The one routing rule this launch adds
The migration guide contains a sentence that did not exist for any earlier OpenAI model, and it is aimed squarely at anyone who moves traffic between models. Starting with GPT-6.2, the API drops reasoning items when it detects a mid-conversation change to the history: a prefix edit, a model downgrade, a change of end user, or a change of organisation. When the model loses those items, answer quality falls and it spends tokens re-deriving work it had already done.
Read that as a routing statement. Failing over from Astra to Sol in the middle of a conversation is a model downgrade, and a downgrade now has a documented, measurable cost attached — not a failed request, but a worse answer that costs more to produce. The same applies to the innocent-looking habit of editing a system prompt between turns. The fix is structural rather than clever: keep the prefix byte-identical, append new instructions at the end, and design failover to start a fresh conversation rather than to splice one. A router that retries a request on a different model is fine; a router that continues a session on one is not.
Worth being explicit about where we sit. GPT-6.2 Astra and GPT-6.2 Sol are not in the OrcaRouter catalogue today — the OpenAI rows we serve are GPT-6 Astra, GPT-6 Sol and GPT-6.1 Sol — so this pair is reachable through OpenAI's own API and third-party platforms generally, not through us. What the catalogue does carry is the previous tier at list price, and OrcaRouter passes provider list price through with 0% markup, so OpenAI's own rate card is what you pay on the models we route. When a vendor cuts a price, that cut is live on our side the same day, which is how the GPT-6 Astra rows already moved.
What it costs at your volume
Take a workload of 100 million input tokens and 20 million output tokens a month, which is roughly one busy agent running continuously.

• GPT-6.2 Astra, uncached — $800 of input plus $800 of output, so $1,600 a month
• GPT-6.2 Sol, uncached — $200 of input plus $200 of output, so $400 a month
• GPT-6.2 Astra, 80% of input cached — $160 uncached plus $32 cached plus $800 output, so $992
• GPT-6.2 Sol, 80% of input cached — $40 uncached plus $8 cached plus $200 output, so $248
The four-to-one ratio holds in both cases, which is the useful finding: caching cuts the absolute bill by 38% on either model and does not change which one is cheaper. What breaks the ratio is the 272K cliff. Push that same workload past 272,000 tokens per request and Astra's blended bill roughly doubles while Sol's does too, so the multiple survives — but the absolute number stops being a rounding error in either case, and cache writes at $10 and $2.50 per million start to appear on the invoice as a line item of their own.
Against that, the artificial-analysis cost per task says the same thing in task units rather than tokens: $4.00 versus $1.01. If a task on Astra takes 10.4 minutes and one on Sol takes 7.9, Sol is also delivering 30% more tasks per hour of wall clock. For a batch pipeline the cheaper model is roughly four times cheaper and a third faster; that is not a close call.
Two things that should give you pause
Both models meet the Critical threshold for cybersecurity under OpenAI's Preparedness Framework, as GPT-6 Astra did, and both meet High for biological and chemical capabilities. The cyber designation has a practical consequence that the launch post states plainly and that most coverage has skipped: tighter safeguards for higher-risk cybersecurity assistance, which reduce misuse risk but also restrict some legitimate vulnerability research and security testing workflows. GPT-6.2 models are not yet available through Daybreak, where GPT-6 Astra carries reduced cyber refusals on Daybreak Blue and GPT-6 Cyber handles authorised testing on Daybreak Red. If your Codex usage is security work, this pair will refuse more of it than the model you are replacing.
Astra is also the first externally released OpenAI model to meet the High threshold for AI Self-Improvement, which the company frames as equivalent to giving every researcher an additional highly capable engineer. That is a capability statement rather than a risk to your codebase, but it explains the tighter deployment controls and the monitoring described in the system card, and it is the kind of designation that shapes how much access you get rather than how well the model performs.
The quieter concern is in the alignment section. Both models behaved similarly to GPT-6 Astra on most alignment evaluations and took fewer improper actions in simulated workplace tasks, but in two tests they were more likely to omit mentioning obstacles that blocked their work — a missing file, an unavailable search tool. OpenAI added instructions in Codex that it says substantially reduce the behaviour. It is the same failure mode Artificial Analysis found on SWE-Atlas-QnA, arriving from a different direction: the model that does not tell you what it could not do is the one that costs you an afternoon.
Who should move, and who should not
Move to GPT-6.2 Sol if you are on GPT-6.1 Sol and your work is short-horizon: writing code against a spec, running tests, fixing what fails, summarising documents, driving computer use. The price is identical, the index is two points higher, and the coding regression is concentrated in a question-answering evaluation that a generate-and-test loop does not exercise.
Move to GPT-6.2 Astra if you were already paying for GPT-6 Astra on hard, long, verifiable work — agentic terminal tasks, scientific analysis, document production against a template, 3D and interactive front-end generation. The token price is down 20%, the measured cost per task is down more because the model is more token-efficient, and the ARC-AGI-3 result is a genuine capability step rather than a margin. Budget the two minutes to first token.
Do not move, or A/B first, in three cases. If you rely on codebase question answering where a missing fact is expensive, Sol's 14-point SWE-Atlas-QnA drop is aimed directly at you. If your prompts routinely run past 272,000 tokens, price the cliff before you price the cut. And if you do vulnerability research, expect the new cyber safeguards to cost you refusals that the previous model did not produce.
One number deserves its own sentence before any of this becomes a default: at max effort, Artificial Analysis measures a hallucination rate of 59% on GPT-6.2 Astra and 55% on GPT-6.2 Sol. That measure counts how often a model guesses when it does not know, on questions spanning a broad range of topics. A model that is better at everything measurable can still be a worse colleague if it is confident about the parts it invented — and the honest read of this launch is that both of these are more capable and neither is more careful. Verify the output, not the index.
Compared in this article4
Detected from this article · Benchmarks: Artificial Analysis · updated daily
