
GPT-6.2 Astra Cuts Token Prices 20%. The Task Bill Falls 36%.
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 90 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 97 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-0711 tok/s
- OrcaNEWOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 88 tok/s
- OrcaNEWOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 133 tok/s
- OrcaNEWOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$4.00 / $20.00 per 1M tokens · 86 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 124 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 67 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 91 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 43 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 192 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 133 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
- OrcaOrca: OrcaCyber Zero 1.52026-10-10$3.00 / $7.50 per 1M tokens · 131 tok/s
- openaiOpenAI: GPT-6.1 Sol2026-09-2955Intelligence
- anthropicAnthropic: Claude Sonnet 5.52026-09-2855Intelligence
- typesafeTypeSafe: Jev 1.132026-09-24$0.04 / $0.00 per 1M tokens · 186 tok/s
OpenAI shipped GPT-6.2 Astra on 17 November 2026 with a headline you can verify in ten seconds: $8 per million input tokens and $40 per million output tokens, against $10 and $50 for GPT-6 Astra. That is a clean 20% cut on both directions of the meter. The number that actually decides whether your invoice moves is a different one, and it is larger. Artificial Analysis measures GPT-6.2 Astra at $4.00 per Intelligence Index task versus $6.24 for GPT-6 Astra — a 36% reduction in what it costs to finish a unit of work, not in what a token costs.
Both numbers are true, and the gap between them is the whole story. A price cut tells you what the meter charges. A cost-per-task figure tells you what the meter charged and how many times it ran. GPT-6.2 Astra charges less per token and burns fewer of them getting to the same place, so the savings compound. Where they stop compounding — long-context requests, cache writes, and the reasoning-item rules — is where this article spends its second half.
The same task, 36% cheaper — and why that is bigger than the sticker
Artificial Analysis scores GPT-6.2 Astra at 62 on its Intelligence Index, third of 87 models tracked, up from 57 and thirteenth for GPT-6 Astra. Its cost figure is computed over the identical evaluation suite, which is what makes the comparison fair: same tasks, same harness, same scoring, only the model changes. Five index points of capability arrived together with a 36% drop in cost per task.
That combination is unusual. Most releases trade one for the other — you pay more per task to get a smarter model, or you accept a weaker model to get a cheaper one. Here both numbers moved the right way against the predecessor, which is what a genuine generation-over-generation efficiency gain looks like rather than a repositioning.
The six figures that matter, side by side. All pricing is OpenAI's published standard-tier rate; index and cost-per-task figures are Artificial Analysis's independent measurements, not OpenAI's.
• Input — GPT-6.2 Astra $8.00 per million tokens vs GPT-6 Astra $10.00
• Output — $40.00 per million tokens vs $50.00
• Cached input — $0.40 per million at a 95% discount vs $1.00 at 90%
• Intelligence Index — 62, ranked #3 of 87 vs 57, ranked #13 of 87
• Cost per Index task — $4.00 vs $6.24
• Context window — 1,050,000 tokens total, of which 922,000 max input and 128,000 max output
Notice the cache line. The discount deepened from 90% to 95% and the base rate it applies to fell from $10 to $8, so a cached input token now costs $0.40 where it used to cost $1.00 — a 60% cut, three times the headline. If your workload is a long system prompt plus a short user turn, you are getting considerably more than 20%.

Where the saving leaks back out
Three mechanisms take a bite out of the 36%, and all three are documented rather than hidden — you just have to read the pricing page and the migration guide before you commit to a context budget.
The 272K cliff
GPT-6.2 Astra has a 1,050,000-token window, but the price changes shape well before you fill it. Once a request's input passes 272,000 tokens, input and cached input bill at double the standard rate and output at 1.5x — and the multiplier applies to the entire request, not just the tokens above the line. A 300,000-token prompt does not pay a premium on the last 28,000 tokens; it pays double on all 300,000.
At $16 per million input, a single 300K call costs $4.80 before the model writes a word, where a moment earlier — at 271,999 tokens — the same work would have cost $2.17. That is the 20% cut handed back in one step, and then some, because the cliff is a discontinuity rather than a slope. Long-document work that used to fit inside the old window needs a deliberate chunking strategy now, not just a larger budget.
Cache writes are a separate line
Reading from cache is cheap. Filling it is not free: cache writes bill at $10 per million tokens, above the standard input rate. A deep, frequently-invalidated prompt cache can cost more to maintain than it saves to read. The economics favour a stable prefix that survives many turns, which is exactly the pattern OpenAI's own migration notes push you toward.
Dropped reasoning items
Starting with this generation, the API discards reasoning items when it detects a mid-conversation change to the history — a prefix edit, a model downgrade, an end-user or organization change. When the model loses that context it can answer worse and spend more tokens re-deriving work it had already done. That second effect is the one that shows up on the bill: an integration that edits its system prompt between turns will quietly pay for the same reasoning twice, and the 36% saving will look like a rounding error.
The fix is architectural rather than financial. Keep the prefix unchanged when you change settings between responses, and the earlier input stays eligible for caching at the same time the reasoning survives. One habit protects both lines of the invoice.
The reasoning settings are a budgeting decision
GPT-6.2 Astra will not accept reasoning.effort set to none. OpenAI's migration guide says to start new integrations at medium, and to map any existing none or minimal setting to low. None of that is surprising for a reasoning-first model, but it does mean a straight model-string swap from a non-reasoning predecessor will produce both a worse answer and a larger bill than the migration plan implies.
Reasoning tokens bill as output. At $40 per million, the difference between low and high on a long agent run is not a rounding error — it is frequently the largest single variable in the cost of the task. The measured $4.00 per Intelligence Index task assumes whatever effort configuration Artificial Analysis used; your number depends on yours. Treat the effort setting as the first cost lever you tune, before you touch anything else.

What the launch actually improved, and what it did not
A price cut is only worth having if the model is at least as good. On the long-horizon and scientific evaluations, GPT-6.2 Astra moved decisively. Terminal-Bench-Science rose from 63.3% on GPT-6 Astra to 78.1%, and Terminal-Bench 4.0 lands at 64.1%. AA-AnalystAgent sits at 62.5%, AA-3D at 1,318 Elo, and Harvey's LAB-AA at 15.0%. For agentic work that runs for many steps and needs to stay on task, that is the profile you want, and it is the profile that produces the token efficiency behind the 36%.
Two places it did not move are worth knowing about before you switch a whole pipeline.
The first is competitive programming. FrontierCode-AA 2.0 puts GPT-6.2 Astra at 45.4%, still behind Claude Fable 5.2 at 48.8%. Astra closed a lot of ground elsewhere but has not taken the lead on hard coding, so a team whose workload is mostly FrontierCode-shaped has less reason to migrate on capability grounds — the price argument is the one that applies.
The second is worse than a stall. On AA-Omniscience, GPT-6.2 Astra's hallucination rate is 59.1%, up from 49.8% on GPT-6 Astra. It is more willing to answer when it does not know. That is a cost problem disguised as a quality problem: a confidently wrong answer you then have to detect, discard, and re-run costs you the tokens twice, and the re-run is where your careful per-task arithmetic breaks. Any workflow where a wrong answer is expensive to catch deserves a verification step before the price cut is treated as a saving.

Getting the price without betting the pipeline on it
GPT-6.2 Astra is available through OpenAI's own API and several third-party platforms. It is not on OrcaRouter's catalogue today, so if you want to call it this week you are going direct to the vendor. Nothing in the pricing analysis above depends on where you call it from — but two things about how you call it are worth planning for now, because they are the same two things that bite teams three weeks into a migration.
The first is failover. A model released yesterday has no track record on your traffic, and the honest advice is to run it beside the model it would replace rather than in place of it. Routing both through a single OpenAI-compatible endpoint, with automatic failover across providers, means a bad day on one upstream does not become a bad day for your users while you are still deciding whether the 36% is real for your workload. OrcaRouter carries GPT-6 Astra, GPT-6 Sol and GPT-6.1 Sol, so the A/B against Astra's predecessor can run on one key with no second contract and no code change on the client side.
The second is price pass-through. OrcaRouter bills at 0% markup — the provider's list price, unchanged. That matters specifically for a release like this one: a vendor price cut lands in your rate the same day it lands in theirs, with no renegotiation and no repricing lag. If GPT-6.2 Astra joins the catalogue, the $8 and $40 above are the numbers you will see, not a marked-up version of them.
What it buys you in the meantime is a clean comparison. Run the same eval set through GPT-6 Astra today, note your real cost per task at your real effort setting, and you have the baseline against which the 36% either holds up or does not. Most teams that skip this step end up measuring the headline instead of their own workload, and discover the 272K cliff the expensive way.
Who should move now, and who should wait
Move now if your workload is long-horizon agentic work, document analysis under roughly 250,000 tokens, or anything scientific and terminal-driven — the Terminal-Bench-Science jump from 63.3% to 78.1% is large enough to change what is feasible, and the token efficiency compounds with the price cut.
Move with care if your prompts routinely run past 272,000 tokens. The cliff does not reward a bigger budget; it punishes one. Chunk, or budget for double.
Wait, or at least A/B first, if your work is heavy on hard competitive coding — FrontierCode-AA 2.0 still favours Claude Fable 5.2 — or if a confident wrong answer is expensive in your domain, where a hallucination rate that rose from 49.8% to 59.1% deserves a measurement of its own before it becomes your default.
And in every case, set your reasoning effort deliberately before you compare invoices. The 20% is on the pricing page and the 36% is on the leaderboard. Neither one is your bill yet.
