GPT-6.2 Astra copies Claude Fable 5.2's $8/$40 card and finishes a task for 28% of Fable's cost. GPT-6.2 Sol keeps $2/$10 and matches Opus 5.5 xhigh for a sixth.
OpenAI cut the price of one model and left the other alone. The bigger change on both is how few tokens they now spend to finish the same work.

Photo by Ryan Hutton on Unsplash
gpt-6.2-astra and gpt-6.2-sol went live in the API, ChatGPT and Codex on November 17. Astra costs $8 per million input tokens and $40 per million output, down from GPT-6 Astra's $10 and $50. Cached input fell further, from $1.00 to $0.40. Sol didn't move at all: $2 in, $0.10 cached, $10 out, the same card as GPT-6.1 Sol, per OpenAI's pricing page.
We expected the story to be the 20% cut. It isn't. Astra's new card is Claude Fable 5.2's card to the cent, with two exceptions that both favour Anthropic, and Astra still finishes an Artificial Analysis task for 28% of what Fable charges at the same score. Sol, whose price didn't change, now scores 57.3 at max effort for $1.01 a task. Opus 5.5 needs xhigh and $6.05 to land in the same place.
Fable 5.2's price list, except for two lines
| Per 1M tokens | GPT-6.2 Astra | GPT-6 Astra | Claude Fable 5.2 | GPT-6.2 Sol |
|---|---|---|---|---|
| Input | $8.00 | $10.00 | $8.00 | $2.00 |
| Cache read | $0.40 | $1.00 | $0.20 | $0.10 |
| Cache write (short) | $10.00 (30 min) | $12.50 | $10.00 (5 min) | $2.50 |
| Cache write (long) | $20.00 (12 h) | $25.00 | $16.00 (1 h) | $5.00 |
| Output | $40.00 | $50.00 | $40.00 | $10.00 |
| Over 272K input | $16 / $0.80 / $60 | $20 / $2 / $75 | no surcharge | $4 / $0.20 / $15 |
| Batch | $4 / $20 | $5 / $25 | $4 / $20 | $1 / $5 |
| Fast (2x) | $16 / $80 | $20 / $100 | $16 / $80 | $4 / $20 |
OpenAI pricing and Anthropic pricing, read November 18, 2026. The long-context row is input / cached / output for the whole request once input passes 272,000 tokens.
Input, output, the short cache write, Batch and Fast all match Fable 5.2 exactly. This is the second time it's happened: GPT-6 Astra launched in September on Fable 5's input, output, cache read and Batch rates, also to the cent. The two lines that differ both go Anthropic's way. Fable charges $0.20 for a cache hit, half of Astra's $0.40, and it bills its full 1M window flat. Astra doubles input and charges 1.5x for output on any request over 272K.
So if both models spent the same tokens, Fable would win. Here's one agent turn priced on each: a 100,000-token prompt with 95% served from cache and 3,000 tokens out. GPT-6 Astra bills $0.295, GPT-6.2 Astra $0.198 and Fable 5.2 $0.179. Push the prompt to 400,000 tokens at 90% cached and Astra jumps to $1.108 against Fable's $0.512, 2.2 times as much. The 20% headline cut is really 33% on a cache-heavy turn, because the cache read fell further than anything else.
They don't spend the same tokens
Artificial Analysis ran both new models at all five effort levels. Its cost per task is what one Intelligence Index task costs at list price, so a verbose model pays for its verbosity here. GPT-6.2 rows are highlighted.
| Model and effort | Index | $ / task | Output tokens (run) |
|---|---|---|---|
| Claude Fable 5.2 max | 64.6 | $26.54 | 275M |
| Claude Fable 5.2 xhigh | 62.2 | $14.83 | 144M |
| GPT-6.2 Astra max | 62.1 | $4.00 | 65.8M |
| Claude Fable 5.2 high | 61.8 | $9.08 | 85.3M |
| GPT-6.2 Astra xhigh | 61.6 | $2.52 | 27.1M |
| GPT-6.2 Astra high | 60.1 | $2.07 | 20.0M |
| Gemini 4 Carbon high | 60.0 | $5.48 | 205M |
| Claude Opus 5.5 max | 59.1 | $10.04 | 218M |
| Claude Opus 5.5 xhigh | 57.4 | $6.05 | 101M |
| GPT-6.2 Sol max | 57.3 | $1.01 | 60.9M |
| GPT-6.2 Astra medium | 57.0 | $1.62 | 12.4M |
| GPT-6 Astra max | 56.9 | $6.24 | 54.3M |
| Claude Fable 5.2 medium | 56.4 | $3.80 | 41.0M |
| GPT-6.2 Sol xhigh | 55.7 | $0.65 | 33.3M |
| GPT-6.2 Sol high | 54.9 | $0.56 | 22.7M |
| GPT-6.1 Sol max | 54.6 | $1.27 | 63.7M |
| GPT-6.2 Sol medium | 52.7 | $0.41 | 14.6M |
| GPT-6.1 Terra max | 51.3 | $0.96 | 99.4M |
Artificial Analysis Intelligence Index v5.0.1 model pages (Astra, Sol, Fable 5.2), read November 18, 2026. Carbon is priced at its $2/$10 card, which runs through November 30. Claude token counts are AA estimates because Anthropic uses a different tokenizer.
Astra at xhigh lands at 61.6, 0.2 points under Fable 5.2 high, for $2.52 against $9.08. Same rate card, 28% of the bill. The output-token column explains it: Fable high wrote about 85 million tokens across the run and Astra xhigh about 27 million. Fable's cheaper cache can't make up a gap that size.
Past Astra max's 62.1, Fable is still the only option. Fable 5.2 max scores 64.6, but it costs $26.54 a task. Astra max gets 62.1 for $4.00, and nothing else from OpenAI scores higher.
Cost per task at the same score
Around 57 points
Around 61 points
Artificial Analysis cost per task. The Carbon December figure is our estimate, double its current $5.48, because its input and output rates double on December 1.
The 57-point band is where most production agents sit. GPT-6.2 Sol gets there for $1.01. Opus 5.5 at xhigh costs six times that, Fable 5.2 on medium nearly four times. It also makes GPT-6 Astra hard to justify. Astra 6.2 on medium matches its max score for $1.62 instead of $6.24, 74% less, and Sol max does it for $1.01.
Picking an effort level
Neither model page states a default effort. The Astra and Sol pages only say the API rejects none, and the migration guide tells new integrations to start at medium. The table gives some better rules.
- Astra max costs 59% more than xhigh for half a point. Use xhigh unless you're chasing that last half point.
- Astra on medium (57.0, $1.62) loses to Sol on max (57.3, $1.01). Below Astra high, Sol is the better buy.
- Sol on xhigh beats GPT-6.1 Sol on max, 55.7 to 54.6, for about 49% less per task. Same card, so this is pure token efficiency.
- If you run GPT-6.1 Terra at max, look at Sol on medium. It scores 52.7 to Terra's 51.3 and costs $0.41 a task to Terra's $0.96, even though its tokens cost twice as much.
Those are averages over AA's task mix. OpenAI's own customers report anywhere from 34.5% fewer output tokens (Harvey) to 62% fewer tokens (Legora) against older models, so measure on your own traffic before you commit to an effort level.
Two new ways to pay less, and one to pay more
Background requests went into public beta in the Responses API on the same day. Accept your answer within two hours and it's 30% off Standard; within eight hours, 40% off, according to the background guide. That puts Sol at $1.20 in and $6.00 out on the 8-hour window, and Astra at $4.80 and $24.00. It covers both 6.2 models, GPT-6.1 Terra and GPT-6 Luna, but not GPT-6 Astra or GPT-6.1 Sol. Batch is still 50% off if 24 hours is fine. The new windows are for jobs that can't wait a day but don't need an answer now.
Ultrafast went the other way. Both 6.2 models have it at 3x Standard through February 28, 2027: $6 / $30 for Sol and $24 / $120 for Astra. From March 1 it becomes 6x, the multiplier GPT-6.1 Sol and GPT-6 Astra already pay. Until then, 6.2 Sol Ultrafast costs half of 6.1 Sol Ultrafast. If you build on it, budget for the doubling in March. Ultrafast is US data residency only.
Neither 6.2 model has a Flex row on the pricing page, so if you use Flex on GPT-6.1 Sol today, Batch or background requests are the closest replacements.
Migration gotchas
reasoning.effort: "none"returns an error on both models. Coming from GPT-6.1 Sol changes nothing, because it didn't acceptnoneeither. Code written for Terra, Luna or GPT-6 Sol withnoneneeds to move tolow.- Context is the same 1.05M as before (922K input plus 128K output). Knowledge cutoffs moved up to May 18, 2026 for Astra and June 1, 2026 for Sol.
- Tool calling is Responses API only on both 6.2 models, as on GPT-6.1 Sol. Chat Completions works for plain generation.
- GPT-6 Astra and GPT-6.1 Sol aren't on the deprecations page, and GPT-6 Astra keeps its $10 / $50 price. Nothing forces you to move, but nobody should be choosing GPT-6 Astra now.
- OpenAI's launch post claims Astra beats Fable 5.2 by 5.5 points on Agents' Last Exam at 5% of the cost, and its chart data matches DataCamp's transcription. These are vendor-run numbers, not independent ones. In that same table, Fable 5.2 still leads FrontierCode 2.0 (49.1 to 44.2) and BenchCAD.
One model ID for most people, a real decision for Fable users
On GPT-6.1 Sol, switch. You pay the same price, get 2.7 more points at max and a cheaper bill per task at every effort from high up, and the only code change is the model string. GPT-6 Astra is even easier to leave: the new one wins on token price, task price and score, and we couldn't find a workload where the old one comes out ahead.
Fable 5.2 users have a harder call. Fable still wins on prompts over 272K tokens and on cache-heavy turns where neither model thinks much, and it has the top score. For ordinary agent work at around 61 points, Astra on xhigh costs a little over a quarter as much.
We priced the previous generation in GPT-6 Astra and GPT-6.1 Sol, and Anthropic's side in Fable 5.2 cost per task by effort. Full cards are on the GPT-6.2 Astra and GPT-6.2 Sol pages, or plug your own token counts into the calculator.
Sources
- OpenAI API pricing - every OpenAI rate in this post, long-context, Batch, Fast and Ultrafast rows
- GPT-6.2 Astra and GPT-6.2 Sol model pages - context, knowledge cutoff, effort, endpoints
- OpenAI changelog and background requests guide - November 17 launch, 2-hour and 8-hour discounts
- OpenAI launch post - cost and token-efficiency claims, customer figures
- Artificial Analysis - Intelligence Index v5.0.1 scores, cost per task and output tokens
- Anthropic pricing - Claude Fable 5.2 and Opus 5.5 rates
- DataCamp - OpenAI's benchmark table, transcribed