Skip links

GPT-6.2 Astra vs Sol: is Astra worth four times the price?

GPT-6.2 Astra in black and blue type over the grey line Worth it over Sol?, above four circles on a line, the first blue.
GPT-6.2 Astra and GPT-6.2 Sol, the top two tiers of OpenAI's GPT-6 range, released on 17 November 2026.

A quarterly marketing review goes wrong quietly. The Google Ads export counts enquiries, the CRM counts qualified leads, and last quarter’s deck called both of them leads. Catching that is the kind of work OpenAI says its two new models are for: GPT-6.2 Astra at the top of the range, and GPT-6.2 Sol at a quarter of the price.

We ran seven marketing jobs in ChatGPT Work: two that set Astra against Sol, and five that set Sol against the model it replaces. This post sets out what each model costs, what the benchmarks and OpenAI’s own system card show, and which one to use.

The answer in brief

  • What they are. OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on 17 November 2026. Astra is the flagship and replaces GPT-6 Astra. Sol is the mid-priced model, replaces GPT-6.1 Sol and is the default in ChatGPT Work.
  • Price. Astra costs $8 per million input tokens and $40 per million output tokens, 20% below GPT-6 Astra. Sol stays at $2 and $10. These are US list prices.
  • Access. ChatGPT Work and Codex on paid plans, the API and Microsoft Foundry. Neither model is in the Free plan or in ordinary ChatGPT chat.
  • Benchmarks. On OpenAI’s own Agents’ Last Exam, Astra scores 78.3% against Claude Fable 5.2’s 72.8%, at about a twentieth of the cost per task. Independent tests are less flattering: Artificial Analysis scores Astra 62 and Sol 57 against Fable 5.2’s 65, and Vals AI ranks both behind Anthropic’s current models overall, with Astra first for science.
  • The saving. Both models finish a job with fewer tokens. Sol at high effort matches GPT-6.1 Sol’s best score for $0.56 a task instead of $1.27.
  • The catch. Both say less, and both are more likely than their predecessors to leave a known problem out of a report. OpenAI’s system card shows Astra doing so about twice as often as GPT-6 Astra.
  • Our tests. Sol matched Astra’s figures whenever the counting rules were written down, and took about twice as long. Against GPT-6.1 Sol it matched the scored results on four of five jobs, missed one requirement in a client report and wrote 2% to 20% fewer words. No run reported a 17-day tracking gap unprompted. Asked about it, Astra found it both times and Sol once.
  • Our view. Start with Sol. Pay for Astra where a side-by-side test on your own work shows it catching errors Sol misses, and keep a person reading the output of both.

What are GPT-6.2 Astra and Sol, and when were they released?

They are the two upper tiers of OpenAI’s GPT-6 range, released together on 17 November 2026. Astra is the flagship, built for long, multi-step work: scientific research, coding, and documents, slide decks and spreadsheets that have to follow a template. Sol is what OpenAI’s launch post calls its “efficient, versatile model for everyday work”.

GPT-6.2 Astra and GPT-6.2 Sol side by side. Sources: OpenAI’s Astra and Sol model pages and its model guide for Work.

If the table extends beyond the screen, scroll sideways to view all columns.

Detail GPT-6.2 Astra GPT-6.2 Sol
Replaces GPT-6.2 AstraGPT-6 Astra GPT-6.2 SolGPT-6.1 Sol
API model ID GPT-6.2 Astragpt-6.2-astra GPT-6.2 Solgpt-6.2-sol
Context window GPT-6.2 Astra1,050,000 tokens GPT-6.2 Sol1,050,000 tokens
Largest input and output GPT-6.2 Astra922,000 and 128,000 tokens GPT-6.2 Sol922,000 and 128,000 tokens
Knowledge cut-off GPT-6.2 Astra18 May 2026 GPT-6.2 Sol1 June 2026
Reads and writes GPT-6.2 AstraText and images in, text out GPT-6.2 SolText and images in, text out
In ChatGPT Work GPT-6.2 AstraChosen in the model picker GPT-6.2 SolThe default for new chats

The specifications barely move from the models they replace. What changes is behaviour. Microsoft’s Foundry announcement says Sol was built for more concise technical writing, more focused coding and debugging, and better judgement about when to call a tool. That concision runs through both models, and it is the source of both the savings and the complaints covered below.

Neither model knows anything recent. For a November product comparison, give the model dated sources or switch on search and check the references it returns. A large context window gives you room for evidence. It does not make the evidence current.

The release came in a crowded fortnight. Anthropic shipped Claude Fable 5.2 on 9 November, and Google opened Gemini 4 Carbon to all paying customers on 10 November.

Why is there no GPT-6.1 Astra?

OpenAI cancelled it before release. The Astra line skips from 6 to 6.2 because the company cancelled GPT-6.1 Astra on 28 September 2026, a day before its DevDay conference and weeks before the model was due in ChatGPT and Codex. GPT-6.2 is the first release in which both tiers move together.

The reason was behaviour, not capability. Saachi Jain, OpenAI’s head of safety systems, told CNBC the model fell short on two things: staying within the scope and permissions it had been given, and telling users accurately what work it had done.

That history is why the reliability section of this post matters more than usual. OpenAI’s system card for the GPT-6.2 models shows one of those two problems has improved while the other has got worse.

Is there a GPT-6.2 Luna or Terra?

No. The release covers the two upper tiers only. OpenAI’s API pricing page and its pricing guide for Work and Codex list no GPT-6.2 Luna or Terra, so GPT-6.1 Terra and GPT-6 Luna remain the cheaper models.

What do GPT-6.2 Astra and Sol cost?

GPT-6.2 Astra costs $8 per million input tokens and $40 per million output tokens in the API, down from $10 and $50 for GPT-6 Astra. GPT-6.2 Sol stays at $2 and $10, the same as GPT-6.1 Sol. Cached input is $0.40 per million tokens on Astra and $0.10 on Sol, so Astra costs four times as much as Sol on each of the three rates.

API list prices in US dollars, at standard short-context rates. Gemini 4 Carbon’s introductory price ends on 30 November 2026. Sources: OpenAI model pages and API pricing page, Anthropic and Google pricing.

If the table extends beyond the screen, scroll sideways to view all columns.

Model Input, per million tokens Output, per million tokens
GPT-6.2 Astra Input, per million tokens$8 Output, per million tokens$40
GPT-6.2 Sol Input, per million tokens$2 Output, per million tokens$10
GPT-6 Astra (previous flagship) Input, per million tokens$10 Output, per million tokens$50
GPT-6.1 Sol (previous Sol) Input, per million tokens$2 Output, per million tokens$10
Claude Fable 5.2 Input, per million tokens$8 Output, per million tokens$40
Claude Opus 5.5 Input, per million tokens$4 Output, per million tokens$20
Gemini 4 Carbon Input, per million tokens$2 introductory, then $4 Output, per million tokens$10 introductory, then $20

These are US dollar list prices before tax. ChatGPT subscriptions in the UK are billed in pounds at a local checkout price, and API usage is billed in dollars.

A reporting job. A job that uses 50,000 input tokens and 5,000 output tokens costs $0.60 on Astra and $0.15 on Sol. Over 100 reports the difference is $45, before tool charges, retries and tax.

That sum is the buying question. If Astra catches an error that would take an analyst ten minutes to find, 45 cents a report is cheap. If both drafts need the same corrections, Sol has the better case.

An agent turn. With 40,000 input tokens, of which 30,000 come from the cache, and 2,000 output tokens, one turn costs about $0.04 on Sol and $0.17 on Astra, by our arithmetic.

Large files cost more. A request with more than 272,000 input tokens moves to a long-context rate for the whole request: input and cache rates double, and output costs 1.5 times as much. One request with 300,000 input tokens and 5,000 output tokens costs $5.10 on Astra and about $1.28 on Sol.

The rest of the rate card:

  • Batch halves the standard rates for work that can wait. OpenAI also lists a beta discount for eligible background requests: 30% off with a two-hour completion window and 40% off with an eight-hour one.
  • Cache writes cost $10 per million tokens on Astra and $2.50 on Sol with the default 30-minute retention, and $20 and $5 with 12-hour retention.
  • Fast mode doubles the standard rates.
  • Ultrafast costs three times the standard rate until 28 February 2027 and six times from 1 March 2027.
  • Regional processing adds 10%. US and EU data residency are supported on Standard, Fast and Batch, and Ultrafast supports US residency only.
  • Web search costs $10 per 1,000 calls plus the tokens it returns.

The price per token is the wrong number to watch, though. Both models need fewer tokens to reach the same result. OpenAI reports 22% fewer output tokens for Astra than for GPT-6 Astra on its GDPval v2 test, and its headline cost claims are all per task, not per token.

Are GPT-6.2 Astra and Sol in ChatGPT?

Yes, on a paid plan, but only inside ChatGPT Work and Codex. Developers get both through the OpenAI API and Microsoft Foundry. Neither model is in the Free plan or in ordinary ChatGPT chat. OpenAI’s Help Centre says chat continues to run on GPT-6 Sol, which our guide to GPT-6 and Intelligent UI covers.

  • Plans. OpenAI’s pricing guide for Work and Codex lists Plus, Pro, Business, Enterprise and Edu, and its Help Centre adds ChatGPT for Healthcare and FedRAMP workspaces. OpenAI’s release notes for Business say both models are in the Work model picker, and in Codex when they are available for your seat. Enterprise and Edu workspaces have the models off until an administrator enables them.
  • Rollout. Both models were in the Work model picker of our UK account on 17 November 2026. The rollout is gradual, so a missing model on day one is not necessarily a fault. Update the ChatGPT desktop app and the Codex CLI first.
  • Default. New Work chats open on Sol, at Light effort on consumer plans and Medium in Enterprise workspaces. Light is the low setting, which scores 41 on the Artificial Analysis index where Medium scores 53, so move the Power slider to Medium before you judge the model on research. Astra is chosen in the model picker.
  • Usage. Work follows the same structure as Codex: an included allowance, then credits. OpenAI estimates 17 to 180 local messages per five hours for Sol on Plus and Standard Business, and 6 to 50 for Astra. These are estimates, not fixed limits, and Pro plans have no five-hour limit.
  • Speed modes. Fast mode uses the allowance twice as quickly. Ultrafast is on the $500 Pro plan for both models and, for the first time, on the $200 Pro plan, starting with Sol. Eligible Enterprise and Edu plans can have it too. It is not available on Plus, the $100 Pro plan or Business, and it draws down the included allowance at six times the standard rate.
  • API. The model names are gpt-6.2-astra and gpt-6.2-sol, through the Responses or Chat Completions API, with Ultrafast available for each. GPT-6.1 Sol stays available, and OpenAI has announced no retirement date for it.
  • Microsoft Foundry. Both models have been generally available since 17 November 2026 at OpenAI’s prices on Global Standard. The EU data zone costs 20% more, which makes Sol $2.40 input and $12 output per million tokens.

UK subscribers pay for ChatGPT plans in pounds at OpenAI’s localised checkout price, so the dollar plan prices above are US prices.

What does OpenAI claim for GPT-6.2 Astra and Sol?

That Astra beats every model it was tested against on long professional work, that Sol now matches the previous flagship, and that both do it cheaply. The main evidence is Agents’ Last Exam, a test of multi-step tasks in real software. Astra scores 5.2 points above GPT-6 Astra at under a third of the cost per task, and 5.5 points above Claude Fable 5.2 at about 5% of the cost.

OpenAI’s Agents’ Last Exam v1.1 results, each model at its best-scoring effort setting on OpenAI’s chart. Source: OpenAI launch post.

If the table extends beyond the screen, scroll sideways to view all columns.

Model and effort Agents’ Last Exam v1.1 Cost per task
GPT-6.2 Astra, xhigh Agents’ Last Exam v1.178.3% Cost per task$2.89
Gemini 4 Carbon, high Agents’ Last Exam v1.177.4% Cost per task$4.60
GPT-6 Astra, max Agents’ Last Exam v1.173.1% Cost per task$8.99
GPT-6.2 Sol, max Agents’ Last Exam v1.173.0% Cost per task$1.32
Claude Fable 5.2, max Agents’ Last Exam v1.172.8% Cost per task$60.05

Read the table with three cautions. OpenAI ran the tests, and its headline compares each model’s best score: Astra’s comes at xhigh, and at max it scores 77.6% for $4.47 a task. The Fable 5.2 figure leaves out nine tasks where Anthropic’s model handed over to a fallback model. And the closest rival is Gemini 4 Carbon, not Fable 5.2: at high effort it sits 0.9 points behind Astra, at $4.60 a task against $2.89.

OpenAI’s other claims follow the same pattern of a moderate gain at a much lower cost, and most come without absolute scores:

  • GDPval v2 (long knowledge-work assignments in sectors including finance and marketing): Astra is 6.2 points above GPT-6 Astra with 22% fewer output tokens, and Sol is 5.3 points above GPT-6.1 Sol.
  • Terminal-Bench-Science 0.1 (research workflows in a terminal): Astra is nine points above GPT-6 Astra at under a third of the cost.
  • FrontierCode 2.0 (Cognition’s test of whether changes to real codebases meet their maintainers’ standards): Sol is within half a point of Astra at about a fifth of the cost, both at max effort.
  • LifeVal (everyday tasks tailored to one person, such as finding a home that fits a budget and a commute): Sol outscores GPT-6 Astra at about a seventh of the cost.
  • Harvey LAB-AA (legal assignments from start to finish): Astra earns full credit on 15.0% of 120 assignments, more than double Fable 5.2’s pass rate. A legal configuration, Astra for Law, reaches 20.8%.
  • BenchCAD 2.1 Core (3D design): Astra scores about 83%, a point behind Fable 5.2 at roughly a ninth of the cost.
  • AA-3D (Artificial Analysis’s blind preference test between 3D experiences built from the same prompt): Astra ranks first on 1,318 Elo, ahead of Fable 5.2 on 1,271 and Sol on 1,259.

OpenAI also says that many of its own staff use Sol as their default model, and that it writes clearly and naturally. The early reviews below do not all agree.

Do independent benchmarks agree?

Partly. Artificial Analysis and Vals AI both put Claude Fable 5.2 ahead overall, and both show GPT-6.2 Astra and Sol costing less per task. Artificial Analysis scores GPT-6.2 Astra at 62 on its Intelligence Index, three points behind Claude Fable 5.2 and two ahead of Gemini 4 Carbon. GPT-6.2 Sol scores 57, two points above GPT-6.1 Sol and level with GPT-6 Astra.

Artificial Analysis Intelligence Index v5.0.1 scores by reasoning effort. The Claude columns are Artificial Analysis’s Default Fallback configurations. Source: Artificial Analysis.

If the table extends beyond the screen, scroll sideways to view all columns.

Reasoning effort GPT-6.2 Astra GPT-6.2 Sol GPT-6.1 Sol Claude Fable 5.2 Claude Opus 5.5
Max GPT-6.2 Astra62 GPT-6.2 Sol57 GPT-6.1 Sol55 Claude Fable 5.265 Claude Opus 5.559
Xhigh GPT-6.2 Astra62 GPT-6.2 Sol56 GPT-6.1 Sol52 Claude Fable 5.262 Claude Opus 5.557
High GPT-6.2 Astra60 GPT-6.2 Sol55 GPT-6.1 Sol50 Claude Fable 5.262 Claude Opus 5.552
Medium GPT-6.2 Astra57 GPT-6.2 Sol53 GPT-6.1 Sol46 Claude Fable 5.256 Claude Opus 5.548
Low GPT-6.2 Astra46 GPT-6.2 Sol41 GPT-6.1 Sol35 Claude Fable 5.254 Claude Opus 5.535

Two things stand out. Astra gains half a point between xhigh and max, from 61.6 to 62.1 before rounding, while its cost per task rises from $2.52 to $4.00. And Sol at high already matches the best score GPT-6.1 Sol could reach at max.

Selected models at their top effort setting on the Artificial Analysis Intelligence Index v5.0.1, with cost and time per index task as Artificial Analysis reports them. Its Gemini 4 Carbon cost uses Google’s rates from 1 December 2026. Source: Artificial Analysis.

If the table extends beyond the screen, scroll sideways to view all columns.

Model and effort Index score Cost per task Minutes per task
Claude Fable 5.2, max Index score65 Cost per task$26.54 Minutes per task41.1
GPT-6.2 Astra, max Index score62 Cost per task$4.00 Minutes per task12.9
Gemini 4 Carbon, high Index score60 Cost per task$7.76 Minutes per task34.5
Claude Opus 5.5, max Index score59 Cost per task$10.04 Minutes per task24.6
GPT-6.2 Sol, max Index score57 Cost per task$1.01 Minutes per task8.0
GPT-6.1 Sol, max Index score55 Cost per task$1.27 Minutes per task23.4

The cost gap is far wider than the score gap. By our arithmetic, Fable 5.2 costs between six and seven times as much per task as Astra for three more points, and Astra costs about four times as much as Sol for five more. Opus 5.5 costs about ten times as much as Sol for two more.

Astra’s low cost comes from writing less: 65.8 million output tokens across the index against Fable’s 275 million. That number cuts both ways. A model that writes a quarter of the tokens may be efficient, or it may be stopping early, which is what several early testers describe below.

For Sol, the size of the saving depends on the setting you choose.

Bar chart: GPT-6.2 Sol at high costs $0.56 per index task; GPT-6.1 Sol at max costs $1.27 for the same score of 55.
Figure 1. GPT-6.2 Sol at high effort matches GPT-6.1 Sol’s best index score for less than half the cost per task. Source: Artificial Analysis Intelligence Index v5.0.1.
Cost and time per index task by reasoning effort, GPT-6.2 Sol against GPT-6.1 Sol. Source: Artificial Analysis Intelligence Index v5.0.1.

If the table extends beyond the screen, scroll sideways to view all columns.

Reasoning effort Cost per task, GPT-6.2 Sol Cost per task, GPT-6.1 Sol Minutes per task, GPT-6.2 Sol Minutes per task, GPT-6.1 Sol
Max Cost per task, GPT-6.2 Sol$1.01 Cost per task, GPT-6.1 Sol$1.27 Minutes per task, GPT-6.2 Sol8.0 Minutes per task, GPT-6.1 Sol23.4
Xhigh Cost per task, GPT-6.2 Sol$0.65 Cost per task, GPT-6.1 Sol$0.76 Minutes per task, GPT-6.2 Sol5.0 Minutes per task, GPT-6.1 Sol9.4
High Cost per task, GPT-6.2 Sol$0.56 Cost per task, GPT-6.1 Sol$0.59 Minutes per task, GPT-6.2 Sol4.3 Minutes per task, GPT-6.1 Sol7.5
Medium Cost per task, GPT-6.2 Sol$0.41 Cost per task, GPT-6.1 Sol$0.36 Minutes per task, GPT-6.2 Sol3.1 Minutes per task, GPT-6.1 Sol4.3
Low Cost per task, GPT-6.2 Sol$0.18 Cost per task, GPT-6.1 Sol$0.16 Minutes per task, GPT-6.2 Sol1.4 Minutes per task, GPT-6.1 Sol1.9

At the same effort setting the new Sol is cheaper only from high upwards: 5% at high, 14% at xhigh and 20% at max, by our arithmetic. At medium and low it costs slightly more per task than GPT-6.1 Sol did, while scoring six or seven points higher.

The large saving comes from dropping a level. GPT-6.2 Sol at high matches GPT-6.1 Sol’s best score of 55 for $0.56 a task instead of $1.27, which is 56% less, and takes 4.3 minutes instead of 23.4. At medium it beats GPT-6.1 Sol at xhigh for $0.41 instead of $0.76. Artificial Analysis re-measures the timings, so expect them to move.

Where do Claude and Gemini still win?

On Vals AI’s boards, Fable 5.2 and Opus 5.5 lead both GPT-6.2 models on finance, legal research and tax, and Gemini 4 Carbon leads both on the overall Vals Index, in third place on 70.45%. Astra’s clear win there is science.

Selected Vals AI results, as percentages except the cost row. Vals AI ran Fable 5.2 at high effort and the other three at max. Source: Vals AI.

If the table extends beyond the screen, scroll sideways to view all columns.

Vals AI test GPT-6.2 Astra GPT-6.2 Sol Claude Fable 5.2 Claude Opus 5.5
Vals Index v2.2 (overall) GPT-6.2 Astra67.66 GPT-6.2 Sol66.06 Claude Fable 5.271.31 Claude Opus 5.570.14
Science Index GPT-6.2 Astra68.44 GPT-6.2 Sol57.76 Claude Fable 5.263.74 Claude Opus 5.558.65
Finance Agent 2.1 GPT-6.2 Astra60.75 GPT-6.2 Sol58.91 Claude Fable 5.269.47 Claude Opus 5.567.85
Legal Research GPT-6.2 Astra42.79 GPT-6.2 Sol37.02 Claude Fable 5.246.15 Claude Opus 5.550.48
Tax Agent GPT-6.2 Astra21.84 GPT-6.2 Sol24.84 Claude Fable 5.232.09 Claude Opus 5.544.54
Index cost per test GPT-6.2 Astra$10.68 GPT-6.2 Sol$3.18 Claude Fable 5.2$17.94 Claude Opus 5.5$24.77

Sol’s own strengths on Vals are technical. It scores 60.61% on Terminal-Bench 4.0, seventh of 68 models and up from 55.05% for GPT-6.1 Sol, and 58.57% on Terminal-Bench-Science, third of 47. On professional services it is mid-table or lower: 34th of 42 on the Legal Index, 27th of 44 on the Finance Index and 50th of 89 on Tax Agent Bench.

Coding is close at the top. On Terminal-Bench 4.0, Fable 5.2 and Opus 5.5 both score 65.15%, Claude Sonnet 5.5 64.14%, and Astra and Gemini 4 Carbon 63.13% each. Astra costs $4.69 a test there and Sol $1.13, against $14.56 for Fable 5.2 and $12.04 for Opus 5.5.

Coding agents are where the new models stall. On the Artificial Analysis Coding Agent Index v2.1, Astra at max is level with GPT-6 Astra, and Sol is three points below GPT-6.1 Sol. Sol’s fall was driven by a 14-point drop on the SWE-Atlas-QnA-AA test, where it left out key facts more often, as reported by Kingy AI and Implicator. A model trained to say less sometimes leaves out the part you needed.

The pattern across all three sources is consistent. The GPT-6.2 models are rarely the best on a given test and almost always the cheapest of the leaders.

What do early testers say?

They are split, and several of the negative reviews are blunt. Most are about Astra, and the complaints share a theme: the model is hard to direct, writes tersely and tends to stop before the job is finished.

None of these is a controlled test, and most come from pre-release access with the tester’s own prompts. Treat them as warnings about what to test. They describe the same behaviour the benchmarks hint at and the system card measures: low token counts, early stops and terse reporting.

OpenAI’s own material points the same way on length. On its HealthBench Professional test, Sol’s answers averaged 1,869 characters against 3,038 for GPT-6.1 Sol, according to the system card, and its developer guide to GPT-6.2 tells teams to revisit any “be concise” instruction in their prompts.

For a marketing team the writing complaint is the one to check first. A model that compresses a campaign brief into shorthand saves tokens and costs an editor’s time.

How reliable are GPT-6.2 Astra and Sol?

More careful about what they do, and less reliable about what they report. OpenAI’s GPT-6.2 system card records one clear improvement and one clear regression, and both bear on the reasons GPT-6.1 Astra was cancelled.

The improvement: in simulated workplace tasks, both new models took improper actions about half as often as their predecessors. High-severity improper actions fell to 0.64% from 1.20% for Astra, and to 0.71% from 1.32% for Sol. That is the scope problem, acting beyond what the user authorised.

The regression is in reporting. OpenAI says both models were more likely than their predecessors to leave out an obstacle that blocked their work, such as a missing file or an unavailable search tool. Table 8 gives the rates OpenAI measured without its Codex developer prompt.

Rates at which each model failed to report a known obstacle in deliberately obstructed tasks, without the Codex developer prompt. Source: OpenAI GPT-6.2 system card, Figures 35 and 36.

If the table extends beyond the screen, scroll sideways to view all columns.

OpenAI internal evaluation (lower is better) GPT-6.2 Astra GPT-6 Astra GPT-6.2 Sol GPT-6.1 Sol
Left out or misrepresented a known coding blocker GPT-6.2 Astra22.28% GPT-6 Astra11.09% GPT-6.2 Sol24.08% GPT-6.1 Sol14.00%
Did not disclose that a search tool was broken GPT-6.2 Astra23.00% GPT-6 Astra7.58% GPT-6.2 Sol9.33% GPT-6.1 Sol2.75%

OpenAI says the developer prompt it ships in Codex cuts these omissions substantially and that it is investigating the cause. With that prompt, Astra’s rates fall to 8.04% and 8.50%, and Sol’s to 8.88% and 3.92%. The prompt is not applied to direct API requests, and OpenAI measured these rates in its own test setups. If you build your own agent, write an instruction to report blockers, then test it with a missing file and a failed tool.

Sol’s everyday conduct slipped too. In OpenAI’s simulation of Codex traffic, laziness flags rose to 5.25% of turns from 3.32% for GPT-6.1 Sol, instruction-following failures to 4.66% from 3.21%, and confirmed misalignment flags of any severity to 6.78% from 5.13%.

Six other findings are worth knowing:

  • Prompt injection. Sol resisted 99.76% of attacks in OpenAI’s benchmark environments and 91.65% in harder, model-generated ones, where GPT-6 Astra resisted 44.98%. Against the largest attacker budget in a multi-turn jailbreak test it held in an estimated 80.9% of cases, against 63.6% for GPT-6.1 Sol. This matters for any agent that reads web pages or emails.
  • Caveats. In targeted tests, both models sometimes dropped a required qualification or stated an unconfirmed detail as fact. Sol passed 87.1% of the compliance cases, against 90.2% for GPT-6 Astra.
  • Refusals and sensitive topics. Sol answered 86.2% of harmless biology questions without refusing, down from 98.2% for GPT-6.1 Sol, and its score on OpenAI’s MentalHealthBench fell to 50.9% from 57.9%.
  • Cybersecurity. Both models meet OpenAI’s Critical threshold, as GPT-6 Astra did, and are High for biological and chemical capability. With safeguards switched off, Astra achieved code execution in 53.5% of trials on the ExploitBench Internal Port test and Sol in 50.5%, against 31.5% for GPT-6 Astra and 21.5% for GPT-6.1 Sol. OpenAI has tightened the default refusals in response and says this will block some legitimate security testing.
  • Self-improvement. Astra is the first model OpenAI has released that meets its High threshold for AI self-improvement, the measure of how much a model can speed up AI research. Sol does not meet it.
  • Outside testing. In the UK AI Security Institute’s simulated scenarios, GPT-6.2 Astra fabricated medical information when it could not reach its sources, and worked round a user’s restriction on internet access when a task was blocked.

For most businesses the cyber and self-improvement ratings are context. The reporting figures are the practical risk: an agent that hits a broken tool and says nothing produces work that looks complete and is not.

What happened when we tested them?

Sol matched Astra on the scored figures in both reporting jobs and took about twice as long. Astra caught duplicate records Sol missed in the job with no written rules. Against GPT-6.1 Sol, the new Sol wrote less and missed one client report requirement.

We ran seven jobs, 19 runs in all, in ChatGPT Work on the web on 17 and 18 November 2026, starting on the day the models launched.

Astra against Sol: two paid media reporting jobs

Both jobs used synthetic data for fictional clients. Every run used a new chat, and both models got the same files and the same prompt.

Test 1 was a short job with no written rules. Three files, one prompt, one answer, scored on eight checks. Each model ran once at Medium effort and once at High. The files held five traps: a campaign with spend but no CRM records, 14 CRM leads with no campaign ID, three duplicate enquiries, a previous report that called two different measures “leads”, and a fourth file the prompt promised but we did not attach.

Test 2 was a harder job with a written brief. Five files and a client brief asking for six deliverables, among them a Q4 budget with four constraints and an email of 120 words at most. Each model ran twice at High effort, and we fixed 18 checks before the first run. The data was messier: UK-format dates, one platform’s costs including VAT, another billed in dollars, a £250 credit with no campaign, test records and repeat enquiries in the CRM, and an empty website sessions file. A second message then changed the brief in a way that made the budget impossible.

GPT-6.2 Astra against GPT-6.2 Sol on two paid media reporting jobs. Times for test 2 are for the first message.

If the table extends beyond the screen, scroll sideways to view all columns.

Test and setting Checks passed, Astra Checks passed, Sol Time worked, Astra Time worked, Sol
Test 1, Medium Checks passed, Astra8 of 8 Checks passed, Sol6 of 8 Time worked, Astra26s Time worked, Sol49s
Test 1, High Checks passed, Astra7 of 8 Checks passed, Sol6 of 8 Time worked, Astra36s Time worked, Sol1m 15s
Test 2, High, run 1 Checks passed, Astra17 of 18 Checks passed, Sol17 of 18 Time worked, Astra41s Time worked, Sol1m 35s
Test 2, High, run 2 Checks passed, Astra17 of 18 Checks passed, Sol17 of 18 Time worked, Astra47s Time worked, Sol1m 41s
  • Every scored figure was right. All four runs of test 1 gave the correct spend, qualified leads and sales by channel. All four runs of test 2 got 28 figures right, each of which needed the VAT, currency, credit, date, duplicate and test-record rules applied correctly.
  • Astra noticed more without being told. In test 1 it removed the three duplicate enquiries in both runs and Sol in neither. The error was 1% of enquiries and touched no cost figure, so both models reached the same budget recommendation. Only Astra at Medium flagged that the old report used “leads” for two measures.
  • Neither model faked a result. All four test 1 answers said the promised file was missing. All four test 2 runs said the sessions file was empty and gave no conversion rate. When the follow-up made a £30,000 budget impossible, all four said so, allocated the £29,625 the rules allowed and asked for a decision on the last £375.
  • Nobody spotted the tracking outage unprompted. In test 2, Meta reported no conversions for 17 days in August while the CRM logged 14 Meta enquiries. No data-issues list mentioned it, which is the one check every run failed. When we then asked each chat to look at the daily conversions, Astra found it both times and Sol once.
  • The emails were usable. All eight ran between 87 and 101 words, in plain English. Sol wrote its emails as draft cards with a Send button, addressed from the signed-in account. Check what you are clicking.

Sol against GPT-6.1 Sol: five marketing jobs

Each model ran each job once at Medium effort. The first three jobs used the same Search Console figures for aiwiz.uk, the last 28 days against the previous 28. The last two were agent tasks on our live site.

  • Search commentary. A summary, the five biggest gainers and losers among pages and among queries, notes on anything missing or ambiguous, and three actions tied to specific rows.
  • Client report. A 600 to 700 word report for a reader who is not an SEO specialist, with eleven stated requirements.
  • Questions email. A reply to eight questions from the site owner. Three can be answered from the data. Five cannot be answered in full, and the prompt did not say so.
  • Link check. Open six aiwiz.uk pages with the model’s web tools and report each one’s title tag, H1, meta description and published date. One of the six did not exist, and the prompt did not say so.
  • Link audit. Find every link in the body of our live GPT-6.1 Sol post, open each one and report whether it works, redirects or is broken. The post has 30 links.
GPT-6.2 Sol against GPT-6.1 Sol on five jobs in ChatGPT Work, one run per model per job at Medium effort.

If the table extends beyond the screen, scroll sideways to view all columns.

Job Check GPT-6.1 Sol GPT-6.2 Sol Words, GPT-6.1 Sol Words, GPT-6.2 Sol
Search commentary CheckMover lists correct, of four GPT-6.1 Sol4 GPT-6.2 Sol4 Words, GPT-6.1 Sol638 Words, GPT-6.2 Sol563
Client report CheckRequirements met, of eleven GPT-6.1 Sol11 GPT-6.2 Sol10 Words, GPT-6.1 Sol617 Words, GPT-6.2 Sol602
Questions email CheckUnanswerable questions flagged, of five GPT-6.1 Sol5 GPT-6.2 Sol5 Words, GPT-6.1 Sol316 Words, GPT-6.2 Sol252
Link check CheckFields correct on the five live pages, of 20 GPT-6.1 Sol20 GPT-6.2 Sol20 Words, GPT-6.1 Sol346 Words, GPT-6.2 Sol326
Link audit CheckBroken links found, of three GPT-6.1 Sol3 GPT-6.2 Sol3 Words, GPT-6.1 Sol263 Words, GPT-6.2 Sol240
  • Neither model invented a figure. Both found the trap in the commentary data, two queries tied for the fifth-largest loss, and both answered the three answerable questions correctly. GPT-6.2 Sol said of all five unanswerable questions that the data could not answer them, which is the reverse of the behaviour OpenAI’s own tests flagged.
  • Both handled a blocked route. In the link check we refused each model’s request to use our Chrome browser. Both carried on in ChatGPT’s cloud browser, returned all 20 fields exactly as they stand in the page source, reported the sixth page as a 404 and invented nothing for it.
  • Both found every broken link. In the link audit both listed all 30 links, found the same three broken ones and a fourth that redirected, and wrongly marked none. GPT-6.2 Sol also noted, unasked, that a link to OpenAI’s migration guide now leads to a page about GPT-6.2. It was the quicker of the two on both agent jobs, by ChatGPT’s own timer: 46 seconds against 56, and 1 minute 2 seconds against 1 minute 29.
  • Sol wrote fewer words on all five jobs. The difference was between 2% and 20%, with the same score on every check but one. Given a word range for the client report it stayed inside it, at 602 words.
  • Its one miss was an instruction. Asked to end the report with exactly three next steps, it gave what reads strictly as four, where GPT-6.1 Sol met all eleven requirements.

We also ran the commentary on GPT-6.2 Sol at Light, the lowest setting on the slider. It matched Medium on every check, in 568 words.

What the tests show

In the job with a written brief, Sol matched Astra on all 28 scored figures and took about twice as long. Astra caught duplicates in the job without written rules, and found the tracking gap more often once we asked. Against GPT-6.1 Sol, the new Sol wrote less and missed one instruction.

We saw none of the compressed prose or early stops that testers describe, and no hidden failure when something was blocked. What both models missed was a problem that sat in the data unannounced.

The tests have limits. We scored them ourselves against answer keys, not blind, and one or two runs per setting cannot separate the models from chance. In the agent jobs every page the models needed could be opened, so no tool failed outright. The obstacles were a missing page and a refused browser, not the missing file or failed search tool that OpenAI tested inside longer coding runs. We did not measure tokens, cost or allowance used, test the API, or use any effort above High.

Should you use GPT-6.2 Astra or Sol?

Sol, for most work. At a quarter of Astra’s token price it now does roughly what the previous flagship did: it matches GPT-6 Astra on Agents’ Last Exam and on the Artificial Analysis index, for $1.01 to $1.32 a task. OpenAI reports it within half a point of Astra on FrontierCode 2.0 at about a fifth of the cost. In our own tests it produced the same figures as Astra whenever the rules were written down.

Astra earns its price in four places:

  • Science and technical research. It beats both Claude models on the Vals Science Index.
  • 3D and visual builds. It ranks first in Artificial Analysis’s blind preference test.
  • Long professional projects of the Agents’ Last Exam kind.
  • Jobs where nobody has written the rules down. In our tests it caught duplicate records Sol missed, and it worked about twice as fast.

Sol has its own weak spots. Its Vals Science Index score is below the previous Sol’s, and it left out more key facts on Artificial Analysis’s code question-and-answer test.

For a marketing team, Sol suits repeatable work where you can give it a clear brief, reliable inputs and a way to check the answer:

  • GA4 and paid media reporting. Supply the exports with the reporting period, attribution settings and conversion definitions, and ask for the commentary and a list of discrepancies. A report nobody reads until morning can run on Batch at half the price.
  • SEO and content research. Supply approved sources and Search Console exports, and ask for a brief that ties each proposed change to a source or a row in the export.
  • WordPress and website QA. Use Codex on a branch to draft a fix, run the existing tests and list anything it could not test. Keep deployment as a separate approval.
  • CRM and marketing operations. Start from a read-only export and ask for proposed field mappings with a reason for each. A team member applies the changes.

Three cases need more care:

  • Long-form writing. Blog drafts, long emails and proposals need a word range and a structure, because both models default to brevity. With both, Sol held the length in our test, though it bent one formatting rule.
  • Agents that browse. If you let ChatGPT Work or a custom agent read competitor pages, inboxes or review sites, the prompt injection results are a good reason to move. Keep a person reading the summary, because the summary is where these models cut corners.
  • Mental health topics. Do not use Sol for customer-facing answers here. The system card shows a lower score than the model it replaces.

The commercial logic is plain. Claude Fable 5.2 scores higher on most independent tests and costs more to run than Astra: about 1.7 times as much on the Vals Index, 3.1 times on Terminal-Bench 4.0 and 6.6 times on the Artificial Analysis index. OpenAI has not closed the quality gap. It has made the gap expensive to care about, provided the model tells you when something went wrong.

So the choice is about the job. Where a wrong answer is expensive and volume is low, the Claude models and Astra earn their price. Where the same job runs hundreds of times a month, Sol does most of the work for a small fraction of the bill.

What should developers check before switching?

Changing the model name is the easy part. Test these on your own workload before a production job moves across.

  1. Start at medium and re-test. OpenAI recommends medium effort for a new integration, and low if you currently use none or minimal. Sol at high already matches GPT-6.1 Sol at max, so test one level below your current setting before you compare bills.
  2. Do not default to max. On the Artificial Analysis index Astra scores 62 at both xhigh and max, and max costs $4.00 a task against $2.52.
  3. Move tool calls to the Responses API. The models accept Chat Completions requests, but tool calling requires Responses. For Chat Completions, remove temperature, top_p, logprobs and top_logprobs. For Responses, remove temperature, top_p and top_logprobs, and take message.output_text.logprobs out of include.
  4. Leave the start of the prompt alone. From GPT-6.2, when you edit earlier turns the Responses API drops the reasoning items that follow the edit. The model may have to repeat that reasoning, which can raise token use and lower quality. Append new instructions instead of rewriting old ones, which also keeps the cache discount.
  5. Write your own blocker instruction. The Codex prompt that reduces omissions is not applied to API requests. Tell the agent to list anything it could not do, and to check that the work is complete before it ends a turn.
  6. Mind the context sizes. Codex gives Sol a 500,000-token context by default, with up to 1 million available through model_context_window. In the API, a request over 272,000 input tokens is billed at the long-context rates.
  7. Check limits and plan for the Ultrafast price change. Sol’s rate limits start at 5,000 requests and 1 million tokens a minute on the Build tier and reach 15,000 requests and 40 million tokens on Grow. Ultrafast doubles in price on 1 March 2027.

Keep gpt-6.1-sol pinned for any workload that fails these checks. Nothing forces a move.

How do you test Astra and Sol on your own work?

On a job you have already signed off, where you know what a correct answer looks like.

  1. Pick one recurring job. Choose one with a clear definition of done, such as a weekly performance commentary whose figures can be reconciled to the source exports.
  2. Run the models side by side. Give Astra, Sol and your current model the same files, tools, instructions and acceptance criteria, starting at Medium effort. Use anonymised or approved data, and make one of the jobs a job with missing information.
  3. Mark the results blind. Have a reviewer score them without knowing which model produced which.
  4. Measure the work through to approval. Record material errors, what each version left out, minutes of human correction, total model and tool charges including retries, and whether it passed first time.
  5. Set the boundaries before connecting live tools. Agree who can grant access, what data may be used and who approves any publication, ad spend, deployment or CRM change.

Three trial jobs suit an agency or an in-house marketing team: reconciling paid media before a budget review, turning an approved case study into a five-slide sales deck on your own slide master, and fixing a WordPress staging bug against written acceptance criteria. For the first, include one campaign with no matching CRM identifier. A good answer flags the gap. A polished table that quietly assigns those leads to a channel fails. We ran two versions of that job ourselves, with the results above.

Four rules apply whichever model wins:

  • Write the counting rules down. In our second test a one-page definitions file was enough for Sol to match Astra on every figure.
  • Ask one direct question about the data. No run reported a 17-day tracking outage unprompted. Asked to look at the daily conversions, Astra found it both times and Sol once.
  • Ask for a list of what it could not do. Have every report end with the files it opened and the figures it could not reconcile, so a reviewer can open a source and check. This is the practical defence against the omission rates in Table 8.
  • Compare before moving finance, tax or legal research. The independent boards put both Claude models ahead on all three.

For the quarterly review at the top of this post, the test is simple. Did Astra catch the difference between enquiries and qualified leads, and did Sol? If both did, use Sol. If Astra keeps catching errors that survive the cheaper workflow, you have found a job where the premium pays for itself.

If you would like help testing GPT-6.2 Astra or Sol against the model you use today, talk to AIWIZ.

Explore
Drag