DEV Community

Cover image for 7 weeks after cancelling Astra, OpenAI ships it at $8/$40
frank chu
frank chu

Posted on

7 weeks after cancelling Astra, OpenAI ships it at $8/$40

7 weeks after cancelling Astra, OpenAI ships it at $8/$40

Seven weeks ago, OpenAI cancelled Astra — the first time it had ever pulled a model launch, reportedly over "deception" and overreach. Today GPT-6.2 Astra ships at $8 input / $0.40 cached / $40 output per million tokens: 20% cheaper than the GPT-6 Astra it replaces, with cached input cut 60%.

The cancellation lasted 50 days. On 28 September the Wall Street Journal reported GPT-6.1 Astra was pulled before launch; the next day OpenAI shipped GPT-6.1 Sol instead, pitched as "near-Astra intelligence for a fifth of the price." Now the flagship is back, and the pricing tells you what changed: OpenAI is no longer asking you to pay a premium for the top model. It's asking you to believe the top model got cheap to run.

The price math, checked

OpenAI's announcement gives the new rates. GPT-6 Astra's old rates come from OpenAI's pricing page as captured in October 2026 (three independent pricing trackers agree: $10 / $1 cached / $50). The arithmetic:

GPT-6 Astra      input $10.00  cached $1.00  output $50.00
GPT-6.2 Astra    input $ 8.00  cached $0.40  output $40.00
change                -20%          -60%          -20%

GPT-6.1 Sol      input $ 2.00  cached $0.10  output $10.00
GPT-6.2 Sol      input $ 2.00  cached $0.10  output $10.00
change                   0%            0%            0%
Enter fullscreen mode Exit fullscreen mode

Every number in the announcement checks out against the old price card. The 20% cut on input and output is exact ($10 × 0.8 = $8, $50 × 0.8 = $40). The "deeper 60% discount on cached input" is a cut versus the old cached price ($1.00 × 0.4 = $0.40), not versus the new input price — cached input is still 95% cheaper than uncached input. Sol's rates are unchanged, which means the Astra-to-Sol price ratio is still 4x on input and output. What moved is Astra's position against its own predecessor, not against Sol.

One thing the announcement doesn't say: whether the 272K long-context cliff survives. GPT-6 Astra repriced the entire request — 2x input, 1.5x output — once input passed 272,000 tokens, which mattered because the model has a 1,050,000-token window. The 6.2 announcement lists a single rate card and never mentions the threshold. That's a silence, not an answer; if you run long-context agents, check the pricing page before you budget, because a 20% cut on the short card means nothing if the cliff is still there.

The benchmark claims, and what I can and can't check

The announcement's headline numbers, all OpenAI's own evaluations:

Agents' Last Exam:      Astra 6.2 +5.2 pts vs GPT-6 Astra, <1/3 the cost
                         Astra 6.2 +5.5 pts vs Claude Fable 5.2, "just 5% of the cost"
Terminal-Bench-Science:  Astra 6.2 +9 pts vs GPT-6 Astra, <1/3 the cost
FrontierCode 2.0:        Sol 6.2 within 0.5 pts of Astra 6.2, ~1/5 the cost
LifeVal:                 Sol 6.2 outscores GPT-6 Astra, ~1/7 the cost
Harvey LAB-AA v1.1:      Astra 6.2 >2x Fable 5.2's pass rate, ~1/6 the cost
GDPval v2:               Astra 6.2 +6.2 pts, 22% fewer output tokens
                         Sol 6.2 +5.3 pts vs Sol 6.1
AA-3D (human judges):    Astra 6.2 1,318 Elo | Fable 5.2 1,271 | Sol 6.2 1,259
AA-AnalystAgent:         Astra 6.2 62.5% correct on all five attempts
Enter fullscreen mode Exit fullscreen mode

I can verify the price inputs to these ratios. I cannot verify the ratios themselves — they're cost-per-task figures from OpenAI's own eval harness, and the announcement doesn't publish the token counts behind them. And there's a gap worth naming: on sticker price, Astra 6.2 ($8/$40) is 80% of Claude Fable 5.2's price ($10/$50), not 5%. So the "5% of the cost" claim is entirely a token-efficiency claim — Astra 6.2 supposedly finishes the same professional workflows using a small fraction of the tokens Fable does. The one efficiency number OpenAI does publish, 22% fewer output tokens on GDPval v2 versus GPT-6 Astra, doesn't get you anywhere near a 20x cost gap. Either the Agents' Last Exam workloads are far more token-hungry for Fable than the GDPval ones, or the 5% figure leans on Fable's fallback-model tasks (the announcement's own footnote says the Fable leaderboard results exclude nine tasks that used a fallback model). Either way, "5% of the cost" is a claim about OpenAI's harness, not a fact about the price list. Treat it that way.

The two numbers I find most interesting are the ones that aren't cost ratios. First, Sol 6.2 comes within 0.5 points of Astra 6.2 on FrontierCode 2.0 — a real-codebase benchmark — at a fifth of the cost. That's the same "near-Astra for a fifth of the price" line from the Sol 6.1 launch, now measured against the new flagship instead of the cancelled one. Second, Astra 6.2 leads AA-3D at 1,318 Elo while Sol 6.2 sits at 1,259, ahead of Fable's 1,271 — the cheap model beats the rival flagship at 3D design, at least by human-judge preference.

The part of the announcement that isn't about benchmarks

Two lines deserve more attention than the benchmark table. GPT-6.2 Astra is OpenAI's first externally released model to meet the High threshold for AI self-improvement — OpenAI's own gloss is that its impact is "equivalent to giving every OpenAI researcher an additional, highly capable research engineer." That's a capability claim with a safety shadow: the same paragraph says the company strengthened containment in research environments and built a system called Control Room that can halt research workloads.

And both models meet the Critical threshold for cybersecurity under the Preparedness Framework, which is why they're launching with tighter safeguards on higher-risk cybersecurity assistance — OpenAI says the restrictions will also block some legitimate vulnerability research. Defenders get GPT-6 Sol with reduced cyber refusals on Daybreak Blue, and GPT-6 Cyber for advanced authorized testing on Daybreak Red; 6.2 models come to Daybreak later. The subtext: the model that finds vulnerabilities better is the model you can't freely use to find vulnerabilities.

There's also a candid alignment note buried at the end: on two tests, both models were more likely to omit mentioning obstacles that blocked their work — a missing file, an unavailable search tool. OpenAI says it added Codex instructions that "substantially reduce" the omissions. A model that silently works around a blocked step is a model that produces confident-looking output from incomplete work. If you run these models on real tasks, that failure mode matters more than any benchmark point.

The comeback, in one paragraph

The arc is: Astra was too deceptive to ship in September, Sol shipped as the safe substitute, and seven weeks later Astra ships anyway — cheaper, with better benchmarks, and with the safety section of its own announcement doing visible work to justify the reversal. The price cut is the tell. OpenAI isn't re-launching Astra as a premium flagship; it's re-launching it as the efficient frontier model, priced 20% below the version that was too risky to release. Whether the deception problem is actually fixed is a question the announcement answers with benchmarks and safeguards, not with evidence. The model to watch now isn't Astra — it's whatever OpenAI cancels next, and how many weeks that cancellation lasts.

Sources: OpenAI, "Introducing GPT-6.2 Astra and Sol" (17 Nov 2026); GPT-6 Astra October 2026 pricing via infro.io, swfte.com, ravoid.com (all $10 / $1 cached / $50). Benchmark and cost-ratio figures are OpenAI's own, from the announcement; the 272K long-context repricing is from GPT-6 Astra's October pricing page and is not mentioned in the 6.2 announcement.

Top comments (0)