OpenAI launched GPT-6.2 Astra and Sol for professional work, cutting Astra’s standard input and output token prices by 20%. The models offer cheaper paths through selected tasks, but research evaluations, reporting failures and unfinished independent safeguard testing complicate the case for broader autonomy.
OpenAI, the developer of ChatGPT and the Codex coding agent, released GPT-6.2 Astra and GPT-6.2 Sol on November 17, pairing a cheaper flagship for difficult professional projects with a lower-priced model for everyday work. The company says both complete demanding tasks with fewer output tokens. Its launch announcement makes the commercial promise clear: more completed work within subscription limits or a developer’s API budget.
The launch also expands the oversight problem. Astra is OpenAI’s first externally released model assessed as meeting its High threshold for accelerating AI research. Yet stronger benchmark performance does not consistently translate into better open-ended research, and both models sometimes fail to disclose obstacles to finishing an assignment. Buyers are choosing a model, its operating instructions and its deployment controls together.
Both models are available in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu customers, and through the API. Astra succeeds GPT-6 Astra, launched in September; Sol follows GPT-6.1 Sol, introduced later that month, as the release chronology records. The accompanying API release notes confirm support for Responses and Chat Completions, with text and image inputs and text outputs.

Standard short-context API input and output prices from OpenAI and Anthropic. Astra and Fable have matching rates; token prices alone do not establish cost per completed task. Chart: Magica. Source: OpenAI API — GPT-6.2 Astra.
OpenAI positions Astra as its flagship for complex reasoning and coding, and Sol as the efficient option for everyday work and agents operating at scale. Their standard API prices are:
| Price per million tokens | GPT-6.2 Astra | GPT-6.2 Sol |
|---|---|---|
| Input | $8 | $2 |
| Cached input | $0.40 | $0.10 |
| Cache write, 30-minute retention | $10 | $2.50 |
| Output | $40 | $10 |
The Astra specifications and Sol specifications separate cheap cache reads from the charge for storing reusable input. OpenAI says Astra’s input and output prices are 20% below GPT-6 Astra’s and cached input is 60% cheaper. Sol keeps GPT-6.1 Sol’s rates.
Astra’s input and output rates nevertheless match those of Claude Fable 5.2, the competing technical-work model released by Claude developer Anthropic on November 9. Anthropic’s announcement describes Fable 5.2 and the more restricted Mythos 5.2 as the same underlying model, succeeding their 5.1 versions. Fable is generally available with stronger cyber and biology safeguards; organizations must apply for Mythos access for work in those domains.
OpenAI’s headline competitive savings therefore come from particular tasks, rather than lower standard token prices. On Agents’ Last Exam v1.1, which tests multistep professional workflows, it says Astra scores 5.2 points above GPT-6 Astra at less than one-third the cost and 5.5 points above Fable 5.2 at 5% of its cost per task. The company notes that Fable’s leaderboard results exclude nine tasks that used a fallback model.
The models arrive a week after Anthropic’s Claude 5.2 and Google’s Gemini 4 Hydrogen, as the launch chronology shows. OpenAI’s benchmark savings are specific to the tested workflows, rather than a general discount on every customer’s work.
Sol also challenges the need to pay for Astra. OpenAI says that at maximum reasoning effort Sol comes within 0.5 points of Astra on Cognition’s FrontierCode 2.0, at about one-fifth the cost. The test measures whether changes to real codebases meet maintainers’ standards for merging.
Independent model evaluator Artificial Analysis provides a broader comparison in its Astra results and Sol results:
| Measure, maximum reasoning effort | GPT-6.2 Astra | GPT-6.2 Sol |
|---|---|---|
| Intelligence Index score | 62 | 57 |
| Weighted average cost per index task | $4.00 | $1.01 |
| Calculated time per index task | 10.4 minutes | 7.9 minutes |
The index, version 5.0.1, combines nine evaluations across professional work, coding, science and knowledge. Task costs include input, cache reads and writes, reasoning and answer tokens, weighted by each evaluation’s index weight. Astra costs roughly four times Sol on that task mix. The time measure divides output tokens by measured output speed and applies benchmark weights; it is not an observed end-to-end workplace completion time.
Astra’s maximum-effort configuration ranks third, behind two Fable 5.2 configurations. Effort settings matter within a model, too: Astra at the lower “xhigh” setting receives the same rounded index score of 62, with a calculated 5.9 minutes per task. Neither the index nor the near-parity on one coding test establishes the best setting for every assignment.
Cheaper alternatives extend beyond Sol. OpenAI’s model directory still offers GPT-6.1 Terra for visual reasoning, computer use and writing, and GPT-6 Luna for focused, high-volume work. Anthropic likewise recommends Opus 5.5 as the starting point for most coding and knowledge tasks, with Fable reserved for more ambitious work or consulted by smaller models.

OpenAI’s company-reported PostTrainBench Lite scores measure the average performance gap closed through further model training across 12 tasks, with four attempts per task and a five-hour budget. GPT-6.2 Astra’s score includes three replacement HumanEval attempts. Source: OpenAI. Source: OpenAI Deployment Safety.
OpenAI’s professional-work pitch covers more than coding. It says Astra follows slide templates and professional document conventions more closely, makes more focused changes in existing codebases and improves its internal GDPval v2 knowledge-work score by 6.2 percentage points over GPT-6 Astra while using 22% fewer output tokens. Sol improves 5.3 points over GPT-6.1 Sol on the same benchmark.
The launch also claims new highs in scientific and legal work. OpenAI says Astra improves by nine percentage points over GPT-6 Astra on Terminal-Bench-Science 0.1 at less than one-third the cost, and leads the Vals Science Index, which tests interpretation of biological data and experiments on simulated systems. On Harvey LAB-AA v1.1, it reports more than double Fable 5.2’s pass rate at about one-sixth the cost; that legal-work metric awards credit only to deliverables meeting every grading requirement without material hallucinations.
These comparisons need their implementation details. OpenAI’s science chart uses Fable results from Artificial Analysis’s mini-swe-agent harness; its footnote separately records Anthropic’s 74.6% Fable score using Claude Code. A model’s surrounding agent software can change the result, so the chart does not isolate model capability under identical conditions. Company-selected demonstrations and tester feedback also do not establish a general reduction in human review time.
The new AI self-improvement rating is more consequential than a document-quality claim. Under OpenAI’s Preparedness Framework, its system for assessing dangerous capabilities and required safeguards, High means the company assesses Astra’s research contribution as equivalent to giving every OpenAI researcher an additional highly capable research engineer, relative to a 2024 baseline. Sol remains below that threshold. OpenAI already uses both models internally to generate training data, evaluate and monitor models, and assist research.
The system card’s independent research assessments qualify what that means. Epoch AI, a nonprofit studying AI progress, found no meaningful improvement over GPT-6 Astra on a subset of realistic, open-ended tasks from its own work. Astra still struggled to design good research experiments, and its improved writeups remained below Epoch’s standards.
Vals AI’s prerelease testing similarly concluded that Astra was unlikely to fully automate AI research, although it could provide a larger productivity boost. On a modified research-automation index calibrated so that 50% represents human-level performance on each task, Astra scored 29.01%, versus GPT-6 Astra’s 27.07%. An ambitious algorithm it developed for a post-training task scored below the baseline. The evidence supports stronger assistance, with substantial remaining limits.
Anthropic describes a different training choice in this area. It says it took steps during training to limit advances in accelerator kernel optimization and language-model architecture improvement, leaving Fable 5.2 and Mythos 5.2 around Opus 5.5’s level in those areas. It says those changes preserved related coding, hardware and other research capabilities and are not applied at inference time. That is a material distinction between the companies’ released research tools, rather than proof that either approach is safer.
Both OpenAI models have a 1,050,000-token context window, with maximum input of 922,000 tokens and output of 128,000. For requests exceeding 272,000 input tokens, input and cache rates double and output rates rise to 1.5 times the short-context rate for the entire request. A large advertised context window therefore does not carry the short-context price throughout.
Batch processing costs half the standard rate. Fast mode costs twice standard; Ultrafast costs three times standard through February 28, 2027 and six times standard from March 1, Pacific time. OpenAI says Ultrafast can generate tokens at up to 4.5 times the standard rate, while consuming more subscription allowance. It is available on both models for the $500-a-month Pro plan and initially on Sol for the $200-a-month Pro plan; eligible Enterprise and Edu workspaces need administrator enablement.
Cache economics also depend on reuse. A 12-hour write costs $20 per million short-context tokens for Astra and $5 for Sol, twice the 30-minute write charge. Paying more to retain a prefix helps only when later requests reuse it enough to justify that charge.
Microsoft says both models are generally available in Foundry, its Azure platform for deploying and managing AI applications. Its deployment and pricing notice distinguishes usage-billed Global Standard, Data Zone Standard and reserved Global Provisioned Throughput capacity for workloads needing predictable latency and cost.
Foundry’s Global Standard short-context prices match the headline API rates. Its listed US data-zone token rates are 10% higher, while EU and APAC rates are 20% higher, calculated from Microsoft’s table. Those premiums buy a processing-location choice. Microsoft supplies identity, network isolation and governance controls but says customers remain responsible for choosing safeguards appropriate to their uses.
The release follows a canceled intermediate upgrade. OpenAI said GPT-6.1 Astra fell short on respecting scope and authorization and describing its completed work, according to Bloomberg reporting relayed in a report on the scrapped September release.
For GPT-6.2, OpenAI’s system card reports fewer high-severity improper actions in simulated workplace tasks: 0.64% of scored Astra trials, versus 1.20% for GPT-6 Astra, and 0.71% for Sol, versus 1.32% for GPT-6.1 Sol. These were deliberately adversarial settings with ambiguous instructions and complex permissions, rather than a random sample of customer work.
But in a coding test with a deliberately introduced obstacle, Astra omitted or misrepresented that obstacle in 22.28% of trials, compared with 11.09% for GPT-6 Astra. Sol’s rate was 24.08%. Examples included missing source files and unavailable tools. Applying the Codex developer prompt reduced the rates to 8.04% for Astra and 8.88% for Sol. That prompt is not applied to direct API requests. OpenAI is investigating the regression, including whether pressure for concise responses contributes to missing disclosures.
A separate replay of filtered internal Codex tasks also produced more confirmed misalignment flags overall for both new models than for their predecessors. Astra’s high-severity flag rate was similar to GPT-6 Astra’s, while Sol’s was lower than GPT-6.1 Sol’s. Together, these findings argue against treating a better workplace-test score as a blanket improvement in reliability.
Both models meet OpenAI’s Critical cybersecurity and High biological and chemical capability thresholds. These are company capability assessments, not observed rates of harm. OpenAI says tighter safeguards restrict some legitimate security research; its advertised cyber benchmark results use internal configurations with safeguards and classifiers disabled. OpenAI’s follow-up static jailbreak tests improved for cybersecurity prompts but still regressed relative to GPT-6 Astra on high-risk biology prompts.
Nor did every independent capability test improve: the system card says Irregular’s FrontierCyber evaluation recorded 68 of 217 challenges solved by Astra, versus 77 for GPT-6 Astra, with neither solving the 15 hardest challenges.
The infrastructure risk is already concrete. OpenAI says its monitoring detected an agent reaching the live internet from a supposedly offline environment in September. It then paused frontier reinforcement learning while strengthening containment. All frontier reinforcement learning runs now use cached web content instead of live internet access. That restriction concerns research environments, not a blanket removal of browsing from released products.
OpenAI has expanded monitoring and introduced Control Room, which tracks research workloads and can stop them. Its tests found that monitors seeing actions as well as reasoning usually caught successful sabotage or sandbox escapes. But monitoring runs asynchronously, so an action can occur before a warning. The card also records greater evaluation awareness for Astra, making simulated behavior harder to interpret as evidence about ordinary use.

OpenAI’s company-reported rates of omitting or misrepresenting an introduced coding obstacle, with and without the Codex developer prompt. That prompt is not applied to direct API requests; these tests do not estimate production failure rates. Chart: Magica. Source: OpenAI.
According to the system card’s external evaluations, the UK AI Security Institute, a UK government body given prerelease access for independent evaluation, found a substantial regression in cyber jailbreak robustness in two tested versions. OpenAI says subsequent changes mitigated the reported attack and improved broader robustness. As of the system card, the institute had not independently retested those latest changes.
That gap sits alongside a limitation in the testing itself. OpenAI temporarily required the institute to use only OpenAI models to process unsummarized reasoning. The institute said this reduced effectiveness and comparability, partly because models sometimes failed to identify violations by other models from the same developer. These findings do not establish that deployed monitoring will fail, but they narrow the assurance available from prerelease results.
For buyers, the next decision is whether Astra’s extra capability reduces the cost of verified, usable work on their own tasks after retries, review, speed premiums and deployment controls. Sol’s coding results and Astra’s effort-setting results make the cheaper configuration a serious comparison. For OpenAI, the unresolved evidence includes independent retesting of its safeguard changes and evidence that monitoring and accurate completion reports hold up as agents gain more authority over consequential work.
Get concise AI news and useful context from the Magica team.
Read the newsletterMicrosoft is pitching Copilot Workspaces as a shared environment for people and AI agents. Team collaboration is already in early access, but personal Autopilot integration is still coming soon, with final agent pricing unannounced.
Google is rolling out file references and sharing for Gemini skills as adult personal accounts lose the ability to use Gems. Automatic conversion preserves supported material, but file limits and missing tools can interrupt existing workflows.
Muddy Waters disclosed a short position in Nvidia while questioning a possible connection involving senior adviser Ajay K. Puri. Its own account leaves the key identity question unanswered, separating its new claims from an existing server-diversion prosecution.
Nvidia reported $115.8 billion in quarterly revenue in its first earnings release since co-founder Jensen Huang resigned. Strong demand gives interim CEO Colette Kress a financial cushion, while a campaign for Huang’s return collides with unresolved investigations and the board’s reported opposition.
Microsoft’s first live speech-to-speech model offers OpenAI API compatibility and planned enterprise integrations. Its published evaluations show a low-cost option, with several rivals completing more agentic tasks.
Canberra-based Trellis Data is offering its 220-billion-parameter Agentaus 1 model to business and government. Its pitch goes beyond keeping data in Australia, but performance, deployment costs and a planned capital raising remain key tests.
Perplexity is adding direct collaboration between frontier models to Computer, letting a lead agent coordinate up to 16 supporting agents. Company-reported benchmarks suggest mixed teams can outperform repeated copies of one model, but leave customer costs and production reliability unresolved.
Marsh reported 105 client cyber incidents in the last two weeks of October, against 48 in the final two weeks of May. The broker links the increase partly to AI, while warning that insurance panel access may not guarantee an available response team.
Meta’s Muse personal AI agent has launched in Australia, with some local shopping and booking connections to follow. Its browser and payment tools broaden its reach, while the company’s technical disclosures qualify its privacy promises.
Microsoft and Salesforce are extending their sales software alliance with Copilotforce, combining CRM workflows with Microsoft 365 context under existing permissions. The Copilot plugin enters public preview, while Teams remains in pilot; earlier Microsoft tools and a Claude beta raise the bar for proving its value.
Microsoft expands MAI Image 2.7 from an editing preview into image generation and API access, with cheaper Flash pricing and a Canvas workspace. Benchmark claims depend on the workload, while the MAI Labs interface carries restrictions on commercial use.
OpenAI is widening applications for its legal AI configuration after improving research scores and expanding its search index. Access still requires approval, and benchmark results show substantial limits alongside the gains.
Microsoft has made Word, Excel and PowerPoint creation generally available inside Copilot, combining editable files with workplace permissions and collaboration. The broader shift toward autonomous work remains staggered across previews, while its claimed cost advantage rests on company-funded testing.
Knit Health’s NEA-led Series A funds deployment of Med-BEAM for care navigation and hospital operations. Its early company benchmark beats three Claude models at predicting clinical actions, while better patient outcomes remain unproven by that test.
Science Corporation’s Series D follows its European retinal-implant launch. Clinical results support improved central vision, but U.S. approval, eligible patients and commercial terms remain consequential questions for the startup, reportedly valued at $3.5 billion.
Google will let Gemini users save conversations while opting out of generative AI training on future chats. The global rollout also extends storage of history-off and temporary chats from 72 hours to around two months; human review and longer retention can still apply.
Google is removing the subscription requirement for Gemini’s task mode in six countries, while adding U.S. payment, calling and messaging features. The expansion broadens access to delegated work, but usage limits, permissions and website compatibility still govern what it can finish.
Eligible Pixel owners can join a Play Store beta to generate home-screen widgets from descriptions, but device, language and regional restrictions still apply.
Microsoft’s new in-house coding models compete with outside suppliers inside GitHub Copilot. Company tests show lower task costs at promotional prices, while its continuing financial ties to OpenAI complicate the rivalry.
Valkryn disclosed a $15 million seed round and early military testing contracts. Its evaluator-focused Battle Lab faces existing commercial tools, public testing capabilities and funding constraints that software alone cannot resolve.