OpenAI Cuts GPT-6.2 Astra's Price 20% While Claiming the Capability Lead
GPT-6.2 Astra's API price drops 20% to $8/$40 per million tokens, matching Claude Fable 5.2, with cached input at $0.40. OpenAI claims a 5.5-point lead on Agents' Last Exam, but testers report mixed results.
OpenAI released GPT-6.2 Astra, its new frontier model, alongside GPT-6.2 Sol for everyday use, and cut Astra's API price 20% at the same time. Astra now runs $8 per million input tokens and $40 per million output tokens, exactly matching Claude Fable 5.2's published rate, with cached input at $0.40. Sol stays at $2/$10.
The capability claim attached to the launch is specific: OpenAI says Astra scores 5.5 points ahead of Fable 5.2 on Agents' Last Exam at 5% of the cost. Early reactions are mixed, with testers reporting Astra trails Fable 5.2 on some benchmarks and that its communication style is confusing. Both things can be true at once, and the gap between them is where your eval budget belongs.
Key Takeaways: What Matters Most?
- GPT-6.2 Astra's API price was cut 20% to $8 input / $40 output per million tokens, matching Claude Fable 5.2's rate.
- Cached input on Astra is $0.40 per million tokens; GPT-6.2 Sol stays at $2/$10 for everyday workloads.
- OpenAI claims Astra scores 5.5 points ahead of Fable 5.2 on Agents' Last Exam at 5% of the cost.
- Early testers report Astra trails Fable 5.2 on some benchmarks and a confusing communication style — the claim needs your own evals.
- OpenAI says Astra is its first model to meet the 'High' self-improvement threshold; both models meet the 'Critical' cyber threshold.
Takeaways like these, every weekday morning. Get the briefing
The price move is the strategy, not a side effect
A 20% cut to $8/$40 is a match play, not a discount play. Anthropic priced Fable 5.2 at $8/$40; OpenAI chose to land on the identical number while claiming the capability lead. That combination, same price, claimed better performance, claimed 5% of the cost on a flagship agentic benchmark, is how a vendor converts a competitor's price into an argument against it.
For buyers, the structure of the frontier market is now three-way and visibly converging on price. This week alone, Google shipped Gemini 4 Neon at $3/$15 with a #3 overall score on the Artificial Analysis Intelligence Index and #1 on its Coding Agent Index. When a mid-tier model can hold near-frontier coding performance at roughly a third of Astra's price, the $8/$40 tier has to justify itself on genuinely hard tasks, not on volume work. That is a new discipline for enterprise buyers who have spent three years defaulting to whatever the newest flagship is.
The cached input rate deserves attention too. At $0.40 per million tokens, agents that re-read large context windows repeatedly, the standard agentic loop, get an order-of-magnitude cheaper on the input side. If your workload is retrieval-heavy and repetitive, cache hit rate may matter more to your bill than headline price.
The claim you should trust least
OpenAI's 5.5-point Agents' Last Exam lead at 5% of the cost is the kind of number that sells evaluations, and it is also the number to verify before acting. Early testers report the opposite on some benchmarks, Astra trailing Fable 5.2, plus a communication style that users find confusing. Benchmark disagreements between vendor claims and community testing are now the norm for frontier releases; the winner changes depending on which suite you weight, and the only suite that pays your invoices is the one you run yourself.
The safety thresholds are the other claim worth noting. OpenAI says Astra is its first model to meet the "High" threshold for AI self-improvement, and both Astra and Sol meet the "Critical" cyber threshold. Whatever your framework for evaluating self-improving models, the first one to hit a named self-improvement tier is a governance line that compliance teams should now have a written answer for, even if the definitions behind the labels remain vendor-defined and hard to audit from outside.
The Sol tier is easy to overlook and probably matters more for most buyers. At $2/$10 it is aimed at everyday workloads, and if your usage is mostly summarization, extraction, routing and light coding, the frontier tier's premium is a cost you can test your way out of this week. The gap between Sol at $2/$10 and Astra at $8/$40 is a 4x input spread, which means misrouting routine work to the frontier model is the single most expensive mistake in most stacks.
What To Do About It
- Run a 3-way bake-off before your next commit. Astra at $8/$40, Fable 5.2 at $8/$40, and Gemini 4 Neon at $3/$15 on your top 20 real tasks, the spread between vendor benchmarks and your workload is where budgets are won.
- Measure your cache hit rate first. At $0.40 cached input, if your agentic loop re-reads context 10+ times per session, restructure prompts to exploit caching before you renegotiate anything.
- Downgrade volume work to Sol or Neon. Anything not genuinely frontier, classify a $2/$10-or-cheaper tier against your current model and move it this week.
- Write down your policy for self-improvement-tier models. Astra is the first OpenAI model to meet the "High" threshold; if you have no documented stance on models that improve themselves, that is your gap to close before it is audited.
The Bottom Line
GPT-6.2 Astra at $8/$40 turns frontier pricing into a commodity: same price as Fable 5.2, a claimed 5.5-point lead on Agents' Last Exam, and cached input at $0.40 that quietly restructures agentic costs. But the mixed early reactions, Astra trailing on some benchmarks, a communication style testers find confusing, mean the claimed lead is a hypothesis until your own evals confirm it. The practical move is a three-way test against Fable 5.2 and Gemini 4 Neon at $3/$15, with volume work pushed down to Sol at $2/$10. Price compression at the frontier is now routine; capability claims are still something you have to buy separately, with your own benchmark suite.
Sources
Frequently Asked Questions
Share this article