OpenAI's GPT-6.2 Astra Tops Leaderboards While Slashing API Costs 20% OpenAI released GPT-6.2 Astra and GPT-6.2 Sol through ChatGPT Work, Codex and the OpenAI API, with Astra cutting cost per task by more than two-thirds versus its predecessor and dropping standard API input and output prices 20%, from GPT-6 Astra's $10 per million input and $50 per million output tokens to $8 and $40. OpenAI's leaderboard places GPT-6.2 Astra first at 1,318 Elo, ahead of Claude Fable 5.2 at 1,271 and GPT-6.2 Sol at 1,259, while Sol came within half a point of Astra on OpenAI's coding evaluation at roughly one-fifth the price and holds GPT-6.1 Sol pricing. Both models ship with tighter safeguards after hitting Critical cyber and High bio/chem thresholds, and Ultrafast mode expands to Pro 200 subscribers for the first time, starting with Sol. OpenAI's GPT-6.2 Astra Tops Leaderboards While Slashing API Costs 20% OpenAI drops GPT-6.2 Astra and Sol into ChatGPT Work, Codex, and the API, promising frontier results at a third the cost per task. - OpenAI launched GPT-6.2 Astra and Sol https://openai.com/index/introducing-gpt-6-2-astra-sol/?utm source=alphasignal , cutting Astra's cost per task by more than two-thirds. - Astra tops OpenAI's leaderboard at 1,318 Elo, ahead of Claude Fable 5.2 at 1,271. - Sol comes within half a point of Astra on a key coding test at roughly one-fifth the price. - Astra's API input and output prices drop 20% versus GPT-6 Astra; Sol holds GPT-6.1 pricing. - Ultrafast mode expands to Pro 200 subscribers for the first time, starting with Sol. - Both models hit Critical cyber and High bio/chem thresholds, shipping with tighter safeguards. OpenAI has released GPT-6.2 Astra and GPT-6.2 Sol through ChatGPT Work, Codex, and the OpenAI API, according to its launch announcement https://openai.com/index/introducing-gpt-6-2-astra-sol/?utm source=alphasignal . Astra targets demanding professional, scientific, coding, and agent workloads. Sol serves as the lower-cost general model that OpenAI says many employees use by default. The company says Astra reaches new highs on its professional and scientific evaluations at less than one-third of its predecessor’s per-task cost. Both models also produce fewer output tokens on demanding tasks. Because API charges and generation latency depend partly on token volume, those reductions could lower operating costs and delay usage-limit caps. The benchmark figures come from OpenAI, so production decisions still require testing on representative workloads. Two models, three delivery paths Access began at launch for ChatGPT Plus, Pro, Enterprise, and Edu customers using ChatGPT Work and Codex. API developers can call gpt-6.2-astra and gpt-6.2-sol , with OpenAI’s low-latency Ultrafast mode available for both model IDs. | Option | Positioning | Ultrafast access | Pricing change | |---|---|---|---| | gpt-6.2-astra | Complex agents, coding, science, professional work, and visual generation | API and Pro 500 at launch | Standard input and output rates fall 20% | | gpt-6.2-sol | General-purpose work and cost-sensitive production workloads | API, Pro 500, and the $200 Pro plan at launch | Unchanged from GPT-6.1 Sol | GPT-6 Astra launched at $10 per million input tokens and $50 per million output tokens. Applying the announced 20% reduction puts GPT-6.2 Astra at $8 per million input tokens and $40 per million output tokens under standard pricing. Teams using Ultrafast should verify its separate pricing and rate limits before estimating production costs. Efficiency leads the benchmark pitch OpenAI’s leaderboard places GPT-6.2 Astra first at 1,318 Elo, followed by Claude Fable 5.2 at 1,271 and GPT-6.2 Sol at 1,259. Elo measures relative performance within a comparison pool, so the scores depend on the tested models, prompts, graders, and evaluation method. | Model | Reported Elo | Gap from Astra | |---|---|---| | GPT-6.2 Astra | 1,318 | 0 | | Claude Fable 5.2 | 1,271 | 47 | | GPT-6.2 Sol | 1,259 | 59 | OpenAI also reports a 5.3-point gain for GPT-6.2 Sol over GPT-6.1 Sol on its cited evaluation. In the company’s agent tests, GPT-6.2 Astra consumed 62% fewer tokens than GPT-6.1 Sol on complex legal tasks. That comparison uses different model tiers and generations, so it cannot isolate Astra’s generation-over-generation improvement. On OpenAI’s coding evaluation, Sol finished within half a point of Astra when both used the highest reasoning-effort setting, which allocates more inference compute before producing an answer. OpenAI puts Sol’s cost at roughly one-fifth of Astra’s in that test. For agent developers, the useful production measures will include total tokens, tool calls, retries, wall-clock latency, and completion quality. Screenshots become design inputs GPT-6.2 Astra can use a reference screenshot to reproduce presentation layouts, colors, typography, and spacing across new slides. OpenAI also shows the model generating websites, 3D scenes, engineering dashboards, and other visual artifacts, positioning stronger spatial and layout reasoning as a core improvement. OpenAI says Astra also edits and tests existing codebases more effectively. Early testers reported cleaner code style, narrower tests, and fewer revision cycles. Those gains could reduce Codex runtime and API spending when agents inspect a repository, modify files, run tests, and repair failures across several iterations. Cyber gains trigger tighter controls OpenAI describes GPT-6.2 Astra as its most capable cybersecurity model to date, citing a 22-point gain over GPT-6 Astra on ExploitBench Internal Port. The contamination-controlled benchmark uses recent vulnerabilities to reduce the chance that a model can answer from memorized training examples. An earlier ExploitBench evaluation covering vulnerabilities from June through August 2026 scored GPT-6 Astra at 39.0% and GPT-6.1 Sol at 5.5%. Applying the reported 22-point gain to Astra’s earlier score yields roughly 61%, contingent on identical datasets and scoring. OpenAI labels the newer benchmark an “Internal Port,” making the 22-point improvement the directly reported figure and 61% an arithmetic estimate. GPT-6.2 Astra and Sol meet OpenAI’s Critical cybersecurity threshold under its Preparedness Framework, the same classification previously assigned to GPT-6 Astra. Both models also meet the framework’s High threshold for biological and chemical capabilities. These labels classify capability bands that trigger stronger deployment safeguards. OpenAI is launching both models with tighter controls for high-risk cybersecurity requests. The company is also working to offer them through Daybreak, its controlled-access path for approved defenders, with fewer refusals for authorized security work. Teams running red-team, vulnerability-research, or incident-response workflows should account for approval requirements and possible refusals in automated pipelines. Production routing needs fresh tests 1. Re-run workload benchmarks. Compare both models on representative prompts and tools while recording quality, output tokens, latency, retries, and total task cost. 2. Test Sol as the default route. Its reported coding performance and lower cost make it a candidate for routine requests, with Astra reserved for tasks that show a measurable quality gain. 3. Validate the new model IDs. Stage gpt-6.2-astra and gpt-6.2-sol behind feature flags, then regression-test structured output, tool calling, context handling, and safety behavior. 4. Measure Ultrafast separately. Lower latency can change concurrency needs and user experience, while its pricing and rate limits may alter the cost calculation. 5. Plan for security refusals. Cybersecurity products may need Daybreak approval, human review, and fallback handling for blocked requests. 6. Evaluate visual output directly. Products that generate slides, sites, dashboards, or 3D scenes should test layout consistency, editable output, accessibility, and rendering across target platforms. Google Cloud availability remains unconfirmed Both model names appeared on the Google Cloud Console Quotas page and in Vertex AI before OpenAI’s announcement. OpenAI already has an agreement to distribute models and Codex workflows through Google Cloud’s Model Garden, making broader Google Cloud availability plausible. The confirmed launch paths described here remain ChatGPT Work, Codex, and the OpenAI API, so procurement plans should treat general Vertex AI access as pending until Google publishes a supported listing.