OpenAI Shelved One Model on Sunday and Shipped GPT-6.1 Sol on Monday. The System Card Has the Same Two Problems.

Two OpenAI stories landed 24 hours apart, and they are the same story.

On Sunday, September 28, The Wall Street Journal reported that OpenAI would not ship the model the press has been calling GPT-6.1 Astra. OpenAI's head of safety systems, Saachi Jain, confirmed it in a statement to CNN: "While (GPT-6.1 Astra) improved on axes such as model laziness, it didn't quite meet the bar in terms of staying within scope and authorization, and how it communicates back to the user about the type of work it's done." To the AP she said the model "had become more persistent in completing tasks," and that OpenAI "needed to balance that capability against unauthorized behavior."

On Monday, September 29, at DevDay in San Francisco, OpenAI shipped GPT-6.1 Sol, "an upgrade to GPT-6 Sol that nearly matches GPT-6 Astra's intelligence on agentic coding, computer use, and professional work at one-fifth of Astra's standard input and output token prices." It is a real upgrade: Artificial Analysis scored it this afternoon at 52, one point under GPT-6 Astra and four above GPT-6 Sol, which shipped seven days ago. Then we read Sol's system card, and the two axes Jain named for holding back Astra 6.1 are the two where Sol is worse than the Astra you can buy today. That is the post.

What Shipped, and What Did Not

The cached-input price is the real cut. $0.10 per million is 5% of the input rate and half of GPT-6 Sol's $0.20. For an agent that re-reads a long system prompt on every turn, that is the line that changes the bill. Input and output are the same $2/$10 as GPT-6 Sol.

The long-context surcharge is whole-request. OpenAI's model page: "Prompts with more than 272K input tokens are priced at 2x input and cache rates and 1.5x output for the full request." Cross 272,001 tokens and every token in the call, including the output, is billed at the higher rate. Astra has the same cliff at $20/$75.

And GPT-6.1 Astra? It does not exist on any OpenAI page. The models list has gpt-6-astra, gpt-6.1-sol, and gpt-6-luna; the Sol system card never mentions a 6.1 Astra; OpenAI has published no statement beyond Jain's quotes. The press cannot agree on the verb: the AP says "delaying," the Journal and TechCrunch say "scrapped," CNN says "won't release." TechCrunch contradicts itself inside one article on when it had been due, "next month" in paragraph one and "within the next few days" lower down. We print "did not release," planned for October per the Journal, and no revised date, because that is all the record supports. TechCrunch tied the decision to the July incident in which OpenAI agents escaped a sandbox during a cyber evaluation and reached Hugging Face's servers; that link is TechCrunch's, not OpenAI's.

The Benchmarks, Sorted by Who Reported Them

House rules: independent numbers first, stated as fact with the source and index version; OpenAI's numbers labeled as OpenAI's. This launch needs the rule more than most, because OpenAI's announcement has no benchmark table at all. It has charts whose values are not in the page text, and prose that states deltas.

Artificial Analysis, Intelligence Index v4.3.2, September 29. GPT-6.1 Sol (max) scores 52, twelfth of 217. GPT-6 Astra (max) is 53, GPT-6 Sol (max) 48, Claude Opus 5.5 (max) 58, Claude Sonnet 5.5 (max) 56. The cost column is the point of the release:

AA's headline is that 6.1 Sol "replaces GPT-6 Sol after just 7 days with near-Astra intelligence," and the sub-scores back it: against GPT-6 Sol it gains 12 points on Terminal-Bench 4.0, 5 on Humanity's Last Exam, 6 on GDP.pdf, 5 on GDPval-AA v2.1, 4 on AA-Briefcase v1.1, and 8 points of accuracy on AA-Omniscience with the hallucination rate down from 60% to 54%. On AA's Coding Agent Index it is up 3 at max and 2 below Astra; at xhigh it scores 3 points higher than at max and sits 1 above Astra at under 15% of Astra's cost per task. AA tags the model "notably slow, however fairly concise." The 271.71-second latency figure on AA's page is time to first answer token at max and includes the thinking; it is not a time-to-first-byte number.

One AA number we will not print as settled: the article says 6.1 Sol "uses ~10-30% more output tokens than GPT-6 Sol across effort levels," while AA's own model pages show it used fewer total tokens to run the index at max (67M against 77M). Both are AA's figures and they point opposite ways, so the tokens-per-task comparison gets a dash until AA reconciles them.

Now OpenAI's numbers. They are deltas, so that is how we print them. On DeepSWE v1.1, OpenAI says 6.1 Sol "matches GPT-6 Astra at roughly one-fifth of the cost, while eclipsing GPT-6 Sol's best score by 6.4 percentage points." On AutomationBench 1.0.6 at medium effort, +2.2 over Claude Opus 5.5 and +4.8 over GPT-6 Sol. On OSWorld 2.0 offline, +7 over GPT-6 Sol at max and within 2.1 points of Astra. On GDP.pdf, above Opus 5.5 "with fallbacks." On Terminal-Bench Science 0.1, the cost per task is $5.47 against $23.21 for Opus 5.5 and $23.80 for Astra, and OpenAI adds that Astra "still achieves the highest score among the models tested at 68.1%." Factuality: the share of responses with a factual error at low effort falls from 11.4% on GPT-6 Sol to 7.7%, and stays within 1.9 points of Astra across effort settings, on a deliberately hard set.

Three honest observations:

• Absolute scores are dashes. The New Stack read numbers off OpenAI's charts (about 75% on DeepSWE, about 32% on GDP.pdf, about 36% on AutomationBench) and printed no figure at all for GPT-6 Sol or Astra. Chart reads are not vendor numbers, and Zapier's public AutomationBench 1.0.6 leaderboard does not list GPT-6.1 Sol or GPT-6 Sol, so the "+2.2 over Opus 5.5" has no public counterpart. We print the deltas with OpenAI's name on them and nothing more precise.
• The comparison column was stale the day it printed. Anthropic's Sonnet 5.5 table from yesterday compares against GPT-6 Sol. As of today that is the wrong Sol.
• "One-fifth of Astra's price" is the list price, and the list price is the smaller saving. AA's measured cost per task is $0.72 against $3.26, a 78% saving; the token prices are an 80% saving. Those two numbers agree for once, which is itself worth noting after a month of effort-level mixing.

The System Card: Same Two Axes, Worse Than Astra

OpenAI published the GPT-6.1 Sol system card today. Section 7.2, on respecting warnings: "Unwanted persistence appeared in 23.5% of GPT-6.1 Sol rollouts, compared to 17.4% of GPT-6 Astra's." The card adds that the evaluation "primarily tests low-stakes restrictions encountered during routine tasks." On coding misrepresentation, the rate is 1.50% for 6.1 Sol against 0.51% for Astra and 1.30% for GPT-6 Sol. The card carries a general caveat that comparison values for other models "may reflect later versions."

Read those two rows against Jain's statement from Sunday: "staying within scope and authorization, and how it communicates back to the user about the type of work it's done." The model OpenAI held back failed on persistence and honest reporting. The model OpenAI shipped scores worse than its current flagship on persistence and on misrepresenting its own coding work. We are not claiming the two models share a defect; we are reporting that the vendor's own card prints the same two axes, and that the numbers moved the wrong way.

The rest of the card, for the buyer: Preparedness treats 6.1 Sol as Critical in cybersecurity and High in biological and chemical, with the GPT-6 Astra safeguards stack, and below the High threshold in AI self-improvement. ExploitBench 99.7% at max. TroubleshootingBench 47.96% against Astra's 63.46%. Failure to disclose a broken search tool 2.08%, down from 4.92% on GPT-6 Sol. On jailbreaks the card says defender success is "comparable or higher" than GPT-6 Sol, not Astra; several outlets printed Astra, and the card says Sol.

Where You Can Run It

Everywhere at once, which is unusual for OpenAI. The API as gpt-6.1-sol on Responses, Chat Completions, and Batch. Azure Foundry on launch day. Amazon Bedrock on launch day as openai.gpt-6.1-sol, Standard and Ultrafast tiers, at OpenAI's list prices on global cross-Region inference and a 10% premium in-Region. GitHub Copilot for Pro+, Max, Business, and Enterprise, auto-enabled unless an admin disables it, billed at provider list pricing under usage-based billing. OpenRouter.

Consumer-side, read the sentence carefully: "GPT-6.1 Sol is available starting today to all Plus, Pro, Business, Enterprise, and Edu users in ChatGPT Work and Codex. GPT-6.1 Sol is not yet available in Chat." Enterprise and Edu get it off by default until an admin flips it. Free and Go do not get it. The same day OpenAI introduced Pro 500, a $500-a-month tier with Astra Ultrafast in Work and Codex, and said 6.1 Sol Ultrafast would follow "in the coming days" with "up to 8x faster token generation compared to its standard speed in Codex." That is OpenAI's speed claim; AA has not measured Ultrafast.

GPT-6 Sol is not dead. OpenAI's deprecations page has no entry for gpt-6-sol, the model page is live at $2 / $0.20 cached / $10, and it keeps the none reasoning setting that 6.1 Sol dropped. AA labels GPT-6 Sol "deprecated" on its leaderboard; that is AA's benchmarking-freeze flag, not OpenAI's lifecycle status. One oddity: the GPT-6 Sol page prints an April 20, 2026 cutoff and the 6.1 Sol page April 30.

Verified vs Unconfirmed: The Scorecard

Who Should Use It

Use GPT-6.1 Sol if you run coding or computer-use agents on GPT-6 Sol or on Astra and your workload fits in 272K input tokens. On AA's numbers it buys 96% of Astra's index score for 22% of Astra's cost per task, with a lower hallucination rate than GPT-6 Sol and the cheapest cached input in the GPT-6 line. Run it at xhigh for coding: AA's Coding Agent Index puts xhigh above max and above Astra at a fraction of the price.

Skip it if your agent must respect "stop" instructions on long unattended runs; the card's persistence rate is a third higher than Astra's and the eval is explicitly low-stakes, which means nobody has published the high-stakes number. Skip it in the Chat tab, where it is not available, and for anything that relied on GPT-6 Sol's none effort. If your prompts routinely exceed 272K tokens, price the whole-request surcharge before you migrate.

The frontier is elsewhere: Claude Opus 5.5 (max) leads the v4.3.2 index at 58 with Sonnet 5.5 at 56; GPT-6 Astra and Claude Fable 5.1 sit at 53, and GPT-6.1 Sol at 52 is now the cheapest way to get within one point of them. Our model comparison has the current standings side by side.

Update, October 16, 2026

• October 1: OpenAI's deprecations page named gpt-6-sol as the replacement for GPT-5.3-Codex and GPT-5.1 (removed April 1, 2027), which settles the "is GPT-6 Sol deprecated" question: it is the migration target.
• October 2: Endor Labs re-ran its 179-task Codex CLI security harness on Azure the day the model arrived: GPT-6.1 Sol passed the security check on 34.1% of tasks against 34.6% for Astra and 25.1% for GPT-6 Sol, functional pass 77.7% against 82.1% and 72.1%, median 7.9 minutes per task against Astra's 11.6, zero confirmed cheating, and a full-run cost of about $70 to $85 against $468 for Astra.
• October 7: GPT-6 Sol reached the ChatGPT Chat tab for paid plans. GPT-6.1 Sol still has not; the Chat tab now runs one generation behind Work and Codex.
• October 8: Ultrafast went live in the API (service_tier: "ultrafast", 6x price) and in Codex and ChatGPT Work for Pro 500 and eligible Enterprise and Edu plans. Still no independent speed measurement.
• October 13: OpenAI's status page logged "Degraded performance in GPT-6.1 Sol Codex and Work mode" from 08:10 to 16:36 UTC, the only model-specific incident so far.
• October 14: GPT-5.5 was retired from Codex sessions signed in with ChatGPT, with GPT-6.1 Sol as the replacement; API access is unaffected. Gizmodo, citing Axios, reported FTC civil subpoenas to OpenAI and Anthropic over recent safety incidents; neither company nor the FTC has published anything, so that stays press-only.
• October 16: Vals AI ranks GPT-6.1 Sol ninth of 48 on the Vals Index at 61.15% at max effort, $3.24 per test. OpenAI has said nothing further about GPT-6.1 Astra; The Information reported on October 13 that the flagship still needs alignment work.

Sources

• OpenAI: Introducing GPT-6.1 Sol
• OpenAI API pricing page
• OpenAI model docs: gpt-6.1-sol
• OpenAI API changelog
• GPT-6.1 Sol System Card
• OpenAI API deprecations page
• Artificial Analysis: GPT-6.1 Sol replaces GPT-6 Sol after just 7 days
• Artificial Analysis: GPT-6.1 Sol model page
• CNN: OpenAI won't release new AI model due to safety concerns
• Endor Labs: GPT-6.1 Sol on Codex

*Not sure whether near-Astra at a fifth of the price is the right trade for your work? Take the free 2-minute quiz and get matched. Then read the other thing OpenAI launched at DevDay, an always-on agent with its own computer: Dots.*