GPT-6.2 Astra vs Qwen4-Max vs Claude Fable 5.2: price, context and benchmark claims
Three flagship models arrived between 9 and 17 November 2026. Two of them carry identical list prices, the third costs under a fifth as much per token, and the benchmark claims each lab published barely overlap. Independent results from Artificial Analysis and the Agents' Last Exam leaderboard fill in some of the gaps.
GPT-6.2 Astra and Claude Fable 5.2 cost the same per token, and Qwen4-Max costs less than a fifth as much. OpenAI’s new flagship, released on 17 November 2026, lists at $8 per million input tokens and $40 per million output tokens. Anthropic’s Fable 5.2, released on 9 November, lists at exactly the same rates. Alibaba’s Qwen4-Max, released on 16 November, lists at $1.50 and $4.50 on its international deployment. All three accept about a million tokens of context and none publishes its weights. The differences that matter sit below the list price: what cached input costs, what happens to the bill above a context threshold, how many tokens each model spends on a task, and which benchmark claims an outside party has checked.
The independent results available on 18 November complicate any single ranking. Artificial Analysis places Fable 5.2 first on its Intelligence Index, GPT-6.2 Astra a few points behind and Qwen4-Max well back. GPT-6.2 Astra leads Artificial Analysis’s run of Terminal-Bench-Science. On the Agents’ Last Exam leaderboard, Fable 5.2 and Qwen4-Max are effectively tied on pass rate, and GPT-6.2 Astra had not yet been added. Fable 5.2 was also, by a wide margin, the most expensive model per task in the Artificial Analysis runs, despite sharing Astra’s price card.
The Frontier Wire has not run any of these models. This comparison reads each vendor’s launch post and developer documentation, Artificial Analysis’s published model pages and evaluation data, and the Agents’ Last Exam leaderboard, all retrieved on 18 November 2026. Vendor benchmark figures are labelled as vendor claims. Where a number is our arithmetic, the text says so. For the wider price landscape across OpenAI, Anthropic, Google, DeepSeek and xAI, see our survey of LLM API pricing; this piece does not repeat it.
TL;DR
- GPT-6.2 Astra and Fable 5.2 both list $8 input and $40 output per million tokens. Fable’s cache reads are cheaper ($0.20 against $0.40), and Fable keeps one price across its whole 1M context while Astra doubles input rates above 272K tokens.
- Qwen4-Max lists $1.50 input, $4.50 output and $0.075 for cache reads on Alibaba Cloud’s international deployment, with one price tier up to 1M tokens.
- Artificial Analysis scores Fable 5.2 at 65, GPT-6.2 Astra at 62 and Qwen4-Max at 43 on its Intelligence Index, at $26.54, $4.00 and $3.27 per task.
- The three labs’ own benchmark charts barely overlap. Alibaba compares Qwen4-Max with older OpenAI and Anthropic models, not with Fable 5.2 or GPT-6.2 Astra.
Criteria for this comparison
Comparisons on this site publish their criteria before any verdict. These are the dimensions used here, in the order they are discussed, with the source of evidence for each.
| Criterion | What is compared | Evidence used |
|---|---|---|
| List price | Standard input and output price per million tokens | Vendor pricing and model pages |
| Cached input | Price of a cache read, cache-write charges and cache lifetimes | Vendor pricing and caching docs |
| Long-context rules | Whether price changes with prompt length | Vendor pricing pages |
| Context and output limits | Context window, maximum input, maximum output | Vendor model pages |
| Availability | Where the model can be called, data residency, consumer plans | Vendor launch posts and docs |
| Licence | Whether weights are published and on what terms | Vendor posts; Artificial Analysis classification |
| Vendor benchmark claims | What each lab reports, against which competitors | Launch posts and embedded chart data |
| Independent results | Third-party scores, cost per task and speed | Artificial Analysis; Agents’ Last Exam leaderboard |
No weighting is applied. Teams running long agent sessions will weight caching and long-context rules; teams on fixed budgets, cost per task; teams moving workloads between clouds, availability.
The three models at a glance
| GPT-6.2 Astra | Qwen4-Max | Claude Fable 5.2 | |
|---|---|---|---|
| Vendor | OpenAI | Alibaba (Qwen team) | Anthropic |
| Release date | 17 November 2026 | 16 November 2026 | 9 November 2026 |
| API model ID | gpt-6.2-astra | qwen4-max | claude-fable-5-2 |
| Input / output price per 1M tokens | $8 / $40 | $1.50 / $4.50 (Singapore, international) | $8 / $40 |
| Cached input per 1M tokens | $0.40 | $0.075 | $0.20 |
| Context window | 1,050,000 tokens | 1,000,000 tokens | 1M tokens |
| Maximum input | 922,000 tokens | 991,808 (983,616 in thinking mode) | Not separately stated |
| Maximum output | 128,000 tokens | 131,072 (262,144 in thinking mode) | 128K tokens |
| Input modalities | Text, image | Text, image, video | Text, image |
| Knowledge cutoff | 18 May 2026 | Not stated | June 2026 (reliable knowledge cutoff) |
| Weights | Proprietary | Proprietary | Proprietary |
Two details are easy to miss. Qwen4-Max is the only one of the three that accepts video input, according to Alibaba Cloud’s model page. And OpenAI caps input at 922,000 tokens inside a 1,050,000-token window, so a prompt cannot fill the whole window.
List prices: one price card shared, one far lower
OpenAI cut GPT-6.2 Astra’s input and output prices 20% from GPT-6 Astra’s $10 and $50, and its cached input 60% from $1, according to its launch post and model pages. Anthropic made the same 20% cut eight days earlier, from Fable 5.1’s $10 and $50: its launch post says Fable 5.2 and Mythos 5.2 are “available at a lower price: $8 per million input tokens and $40 per million output tokens.” The two US labs’ top general-release models now share a list price to the dollar.
Alibaba’s pricing is in a different band. Alibaba Cloud Model Studio lists Qwen4-Max at $1.50 input and $4.50 output per million tokens on its international (Singapore) deployment, for any request up to 1M tokens, with a one-million-token free quota; QwenCloud shows the same prices. Model Studio also lists $1.413 and $4.24 under its other regional tabs. Reasoning tokens are billed as output, as they are for the other two models. By our arithmetic, Qwen4-Max’s input costs 18.75% of the US models’ and its output 11.25%.
Batch processing costs half of standard rates for GPT-6.2 Astra and Fable 5.2 ($4 input and $20 output for Fable). Alibaba’s capability tables list batch inference for Qwen4-Max’s Chinese mainland deployment only. Both US models also sell speed: OpenAI’s Fast mode costs twice Standard and its Ultrafast tier three times Standard through 28 February 2027, then six times from 1 March 2027. Anthropic’s fast mode for Fable 5.2 costs $16 input and $80 output.
Cached input: where the bills diverge first
For agent workloads that resend a long, stable prefix (a system prompt, tool definitions, a codebase or a document set), the cached-input price can matter more than the headline input price. All three vendors now charge separately for writing to the cache and reading from it, so both numbers belong in the comparison.
| Caching term | GPT-6.2 Astra | Qwen4-Max (Singapore) | Claude Fable 5.2 |
|---|---|---|---|
| Cache read per 1M tokens | $0.40 | $0.075 (implicit or explicit) | $0.20 |
| Cache read as share of input price | 5% | 5% | 2.5% |
| Standard cache write per 1M tokens | $10 (30-minute retention) | $1.875 (5-minute TTL) | $10 (5-minute cache) |
| Longer cache write per 1M tokens | $20 (12-hour, short-context Standard) | $3 (1-hour TTL) | $16 (1-hour cache) |
| Minimum cacheable prefix | Not stated on model page | 1,024 tokens | Not covered here |
Sources: OpenAI’s GPT-6.2 Astra model page; Alibaba Cloud’s model pricing page, Qwen4 explicit-cache page and QwenCloud model page; Anthropic’s pricing page.
The cache-read numbers produce a clean ordering. Fable 5.2 charges 2.5% of its input price for a cache hit, which Anthropic’s pricing page says applies to Fable 5.2, Mythos 5.2 and their 5.1 predecessors. GPT-6.2 Astra charges 5%. Qwen4-Max also charges 5% of its own input price, which is $0.075 in absolute terms, less than half of Fable’s $0.20.
Write charges run the other way. OpenAI’s cache write for GPT-6.2 Astra is $10 per million tokens for 30 minutes of retention, which is 1.25 times its input price. Anthropic’s 5-minute cache write for Fable 5.2 is also $10, 1.25 times input, and its 1-hour write is $16, twice input. Alibaba’s explicit-cache page describes the same multipliers for Qwen4: “approximately 125%” of the input price for a 5-minute write and “approximately 200%” for a 1-hour write. Alibaba’s five-minute and one-hour lifetimes, and its 1.25 times and 2 times write multipliers, mirror Anthropic’s published structure. Alibaba’s page adds a renewal rule of its own: a one-hour request that hits a valid five-minute cache is billed as a new one-hour write, not as a read.
The practical difference is lifetime. OpenAI’s base cache write buys 30 minutes, six times the default lifetime on the other two, and its longest option is 12 hours. Anthropic and Alibaba top out at one hour. For a workload where the same prefix is reused every 10 to 20 minutes, OpenAI’s default write would still be live while the others’ 5-minute writes would have expired and need rewriting or upgrading to the more expensive one-hour tier.
OpenAI’s migration guide adds a behavioural condition. Starting with GPT-6.2, it says, “the API will drop reasoning items when it detects mid-conversation changes to the conversation history (prefix edits), model downgrade, end user, or organization,” and it recommends keeping the prefix unchanged so that earlier input stays eligible for caching. Teams that rewrite earlier turns will lose cache hits and reasoning state together.
Long context: one model changes price, two do not
The biggest structural difference in the price cards concerns long prompts. OpenAI’s GPT-6.2 Astra model page states: “For requests with more than 272K input tokens, input and cache rates double and output costs 1.5 times the short-context rate for the full request.” The surcharge applies to the whole request, not only to tokens above the threshold. GPT-6 Astra carried the same rule.
Anthropic’s pricing page says Claude 4.6 and later models, other than Haiku 5.5, “include the full 1M token context window at standard pricing,” with the example that “a 900k-token request is billed at the same per-token rate as a 9k-token request.” Fable 5.2 falls under that rule.
Alibaba’s Qwen4-Max pricing has a single tier, “0<Token≤1M,” so prompt length does not change the per-token rate either. Model Studio uses tiered pricing for some other models, where every token in a request is billed at the rate of the tier the request falls into, but Qwen4-Max is not one of them.
The following worked example shows how these rules play out. It uses list prices, assumes the cache is already warm (no write charges), and ignores tool fees. It is our arithmetic, not a vendor figure.
| Scenario | GPT-6.2 Astra | Qwen4-Max (Singapore) | Claude Fable 5.2 |
|---|---|---|---|
| A: 200K input (150K cached), 10K output | $0.40 + $0.06 + $0.40 = $0.86 | $0.075 + $0.011 + $0.045 = $0.13 | $0.40 + $0.03 + $0.40 = $0.83 |
| B: 500K input (400K cached), 20K output | $1.60 + $0.32 + $1.20 = $3.12 | $0.15 + $0.03 + $0.09 = $0.27 | $0.80 + $0.08 + $0.80 = $1.68 |
In scenario A, below OpenAI’s threshold, the two US models cost almost the same per request, with Fable slightly cheaper because of its cache-read price. In scenario B, the 500K prompt pushes the whole Astra request into long-context pricing: uncached input at $16 per million, cached input at $0.80 and output at $60. Astra’s request then costs nearly twice Fable’s. Qwen4-Max is cheapest in both cases by a wide margin.
Per-request arithmetic is only the first step: a model that needs fewer tokens can cost less per task at higher rates, as the independent results below show.
Availability and access
OpenAI’s launch post says GPT-6.2 Astra reached “all Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex” on 17 November (Enterprise and Edu once an admin enables it) and the API the same day. The migration guide lists US and EU data residency for Standard, Fast and Batch processing, US only for Ultrafast, and Zero Data Retention for approved organisations. Tool calling requires the Responses API. The documents read for this article do not mention third-party cloud availability.
Fable 5.2 has the broadest cloud footprint of the three. Anthropic’s launch post says it is “available today on all platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure,” and the models overview lists IDs for Amazon Bedrock, Claude Platform on AWS, Claude Models on Google Cloud (both Anthropic-managed and Google-managed) and Microsoft Foundry. Anthropic charges 1.1 times standard rates for US-only inference requested through the inference_geo parameter; Bedrock and Google-managed deployments set their own regional prices.
Qwen4-Max is served by Alibaba. The Qwen blog says it is available on QwenCloud through the Token Plan or the API, and in QwenWork, Qoder and QwenStudio. Alibaba Cloud’s model page lists deployments in China (Beijing), Singapore, Germany (Frankfurt), the US (Virginia), Japan (Tokyo) and Hong Kong, and says the model is compatible with the OpenAI Chat Completions and Responses APIs, the Anthropic Messages API and Alibaba’s DashScope API. Fine-tuning and model deployment are listed as not supported. Alibaba’s enterprise post of 17 November tells enterprise buyers that “availability, features and pricing vary by region.”
The safeguard design affects what a user actually receives. Anthropic says Fable 5.2’s classifiers block high-risk cyber activities and most research-biology tasks, and that a blocked query is answered instead by Claude Opus 4.8 for cybersecurity and Opus 5 for biology. The unrestricted configuration, Mythos 5.2, is limited to vetted organisations. OpenAI says GPT-6.2 Astra and Sol meet its Critical cybersecurity threshold and launched “with tighter safeguards for higher-risk cybersecurity assistance,” and that it “will work to bring GPT‑6.2 models to Daybreak,” its vetted cyber-access programme. Alibaba says it filtered offensive cyber material out of Qwen4-Max’s training data. Our survey of how China’s AI labs gate cyber capability covers Alibaba’s approach in detail, and an earlier Frontier Wire comparison of gated cyber models covers Mythos and OpenAI’s Daybreak.
Licence: three proprietary models
None of the three models has published weights. Artificial Analysis classifies all three as proprietary. OpenAI and Anthropic do not release weights for their flagship models.
Alibaba’s position needs more care because the Qwen family includes many open-weight models. Alibaba’s 16 November announcement says that “open-weight, smaller-scale models from the Qwen4 series will soon be released to developers worldwide,” and the Qwen blog says “smaller models in the family are coming to open source.” Neither names a date, sizes or a licence. What Alibaba has open-sourced with this launch is the Qwen-Live-Delivery-Plugin, which the blog describes as a way to install Live Delivery, Alibaba’s term for agents that ship a usable first version of a long task early and then deliver improved versions.
A team that needs to self-host or fine-tune cannot do it with any of the three today: Alibaba and OpenAI both list fine-tuning as unsupported for these models.
What each lab claims, and against whom
The three labs’ benchmark claims rarely measure the same thing on the same models.
OpenAI’s claims. The launch post says GPT-6.2 Astra “sets a new high” on Agents’ Last Exam, “scoring 5.2 points higher than GPT‑6 Astra at less than a third of the cost, and 5.5 points ahead of Claude Fable 5.2 at just 5% of the cost.” It calls the model “the best model in the world for scientists,” citing a new high on Terminal-Bench-Science 0.1 (9 points above GPT-6 Astra) and on the Vals Science Index. The chart data embedded in the post gives the underlying numbers. On Agents’ Last Exam v1.1, OpenAI plots GPT-6.2 Astra at 78.3% at Xhigh effort for $2.89 per task, and Fable 5.2 at 72.8% at maximum effort for $60.05 per task. On Terminal-Bench-Science, it plots GPT-6.2 Astra at 77.1% at maximum effort and Fable 5.2 at 73.3%, with the Fable figure labelled as run on the “mini-swe-agent harness.” On Vals Science Index v1, it plots GPT-6.2 Astra at 68.4% at maximum effort and Fable 5.2 at 63.7% at high effort.
OpenAI’s own charts are not uniformly favourable to its model. On FrontierCode 2.0, the chart shows Fable 5.2 at 49.1% at Xhigh effort, above GPT-6.2 Astra’s best of 44.2%. On the Artificial Analysis Intelligence Index chart that OpenAI reproduced, Fable 5.2 at maximum effort sits at 64.6, above GPT-6.2 Astra’s 62.1. OpenAI’s argument is about cost: at those points, its chart shows GPT-6.2 Astra at about $4.00 per index task against $26.54 for Fable.
Anthropic’s claims. Anthropic’s table compares Fable 5.2 with Fable 5.1, Opus 5.5, GPT-6 Astra and Gemini 4 Carbon, all of which predate GPT-6.2 Astra and Qwen4-Max. It reports Fable 5.2 at 49.1% on FrontierCode 2.0 at Xhigh effort, 69.9% on FrontierSWE v2, 74.6% on Terminal-Bench-Science 0.1 (75.3% for Mythos 5.2), 89.3% on OSWorld 2.1 (offline set, partial), 74.2% on Humanity’s Last Exam with tools, 73.4% on Chartography without tools, and 1785 Elo on AA-Briefcase v1.1. Anthropic says scores show “each model’s highest-available effort” and that Fable figures were run “with safeguards on unless otherwise noted.”
Alibaba’s claims. The Qwen blog’s comparison chart sets Qwen4-Max against Qwen3.8-Max, Claude Opus 5.5, GPT-6.1 Sol, GPT-6 Astra and Kimi K3.1. Neither Fable 5.2 nor GPT-6.2 Astra appears. In that chart, Qwen4-Max scores 46.7 on Agents’ Last Exam v1.1, 75.7 on DeepSWE v1.1, 48.5 on Terminal-Bench 4.0, 51.7 on Humanity’s Last Exam, 54.2 on Chartography, 21.3 on HLAB-Vals-1.1 and 1686 Elo on AA-Briefcase v1.1. Its footnotes say several scores are cited from external leaderboards: AA-Briefcase and AutomationBench-AA from Artificial Analysis, Agents’ Last Exam from its leaderboard, HLAB-Vals from the Vals Index and APEX-Agents from Mercor. Alibaba’s headline claim is about its own predecessor: on Work360, its internal benchmark, Qwen4-Max scored 52.69 against 47.74 for Qwen3.8-Max.
Alibaba’s three launch documents state the predecessor cost saving differently. The Qwen blog says average Work360 cost per task drops from $4.43 to $1.67, a 62% reduction. Alibaba’s 16 November announcement says “around 60%,” and its 17 November enterprise post says 58%. The documents do not explain the difference.
Where the three vendors’ claims touch the same benchmark, the table below collects them. Every cell is a vendor claim, and the conditions differ.
| Benchmark | GPT-6.2 Astra (OpenAI) | Qwen4-Max (Alibaba) | Claude Fable 5.2 (Anthropic) |
|---|---|---|---|
| Agents’ Last Exam v1.1 | 78.3% (chart axis labelled “Score”), Xhigh | 46.7 (pass rate, cited from leaderboard) | Not reported |
| Terminal-Bench-Science 0.1 | 77.1%, Max | Not reported | 74.6%, highest effort |
| FrontierCode 2.0 | 44.2%, Max | Not reported | 49.1%, Xhigh |
| Humanity’s Last Exam | Not reported | 51.7 (conditions not stated) | 74.2% with tools |
| Chartography | Not reported | 54.2 | 73.4% without tools |
| AA-Briefcase v1.1 | Not reported | 1686 Elo (cited from AA) | 1785 Elo |
The Agents’ Last Exam row shows why a cell-by-cell reading misleads. The leaderboard publishes two numbers: a pass rate, “the share of runs that earned a perfect score,” and a score, “the average partial-credit score.” Alibaba’s 46.7 matches Qwen4-Max’s leaderboard pass rate. OpenAI’s 78.3% is consistent with the partial-credit score: on the leaderboard, GPT-6 Astra scores 73.1% and Fable 5.2 72.8%, and adding OpenAI’s stated margins of 5.2 and 5.5 points gives 78.3% in both cases. That reconciliation is our arithmetic. OpenAI’s launch note itself says that Agents’ Last Exam “leaderboard results for Fable 5.2 exclude nine tasks that used a fallback model.”
Independent results: Artificial Analysis
Artificial Analysis runs its own evaluations and publishes model pages with an Intelligence Index, cost per task, speed and token use. Its Intelligence Index v5.0.1 combines nine evaluations: AA-Briefcase v1.1, AutomationBench-AA, Terminal-Bench 4.0, FrontierCode-AA 2.0, AA-Omniscience v1.1, GDP.pdf, HLE-Diamond, CritPt v2 (Hard Subset) and Terminal-Bench-Science 0.1. The figures below are as retrieved on 18 November 2026, for the configuration each page names. Artificial Analysis updates these pages, so they may change.
| Artificial Analysis metric | GPT-6.2 Astra (Max) | Qwen4 Max (Xhigh) | Claude Fable 5.2 (Max, default fallback) |
|---|---|---|---|
| Intelligence Index v5.0.1 | 62 | 43 | 65 |
| Rank among 91 models | 3 | 47 | 1 |
| Cost per Intelligence Index task | $4.00 | $3.27 | $26.54 |
| Time per index task | 10.4 minutes | 31.9 minutes | 40.8 minutes |
| Output speed | 66 tokens/s | 55 tokens/s | 112 tokens/s |
| Time to first answer token | 134.86 s | 147.96 s | 221.15 s |
| Output tokens used on the index | 65.8M | 127M | 275M |
| Input tokens used on the index | 2.16B | 9.9B | 28.3B |
| Cache discount | 95% | 95% | 98% |
| Terminal-Bench-Science 0.1 | 78.1% | 23.3% | 73.3% |
| HLE-Diamond | 72.2% | 44.9% | 80.2% |
| AA-Briefcase Elo | 1644 | 1686 | 1785 |
Several findings stand out.
Identical prices, very different bills. GPT-6.2 Astra and Fable 5.2 have the same price card, yet Fable cost about 6.6 times as much per index task ($26.54 against $4.00, our arithmetic). The token columns explain it: Fable generated about four times as many output tokens (275M against 65.8M) and consumed about 13 times as many input tokens. Artificial Analysis describes Fable 5.2 at maximum effort as “particularly expensive when comparing to other models of similar price” and “very verbose.” Some of Fable’s extra input is likely to have been cached, which its 98% cache discount would soften, but the page does not split cached from uncached input.
Fable leads the index, Astra leads science in the terminal. Fable’s 65 is the highest Intelligence Index score Artificial Analysis lists, and it leads HLE-Diamond and AA-Briefcase among the three. GPT-6.2 Astra leads Terminal-Bench-Science at 78.1% at maximum effort. Artificial Analysis’s evaluation page says GPT-6.2 Astra at Xhigh effort scores slightly higher, 78.6%. That benchmark has 70 tasks, which Artificial Analysis runs with the mini-swe-agent harness and reports as pass@1 averaged over three repeats. The 4.8-point gap between Astra and Fable on it amounts to a few tasks.
Qwen4-Max’s independent numbers are much lower than its positioning. At 43, Qwen4 Max sits at the median Artificial Analysis reports for comparable models and 22 points behind Fable. Its Terminal-Bench-Science result, 23.3%, is less than a third of the US models’ results. Its AA-Briefcase Elo of 1686 matches the figure Alibaba cited and sits above GPT-6.2 Astra’s 1644. Artificial Analysis calls the model “below average in intelligence, but well priced when comparing to other models of similar price,” and “notably slow per task.” At $3.27 per index task, its cost per task is close to GPT-6.2 Astra’s $4.00 even though its per-token prices are a fraction of Astra’s, because it spent nearly twice as many output tokens and over four times as many input tokens.
A vendor-cited score can be one of several. Alibaba’s chart gives Qwen4-Max 72.5 on “Automationbench-AA.” Artificial Analysis’s data for Qwen4 Max lists a partial score of 72.6% and a strict score of 47.2% on AutomationBench. Alibaba’s chart figure is close to the partial score; the chart does not say which scoring was used. Anthropic’s 51.4% for Fable 5.2 on AutomationBench, and OpenAI’s chart figure of 52.4% for GPT-6.2 Astra on AutomationBench 1.0.6, are not labelled with a scoring variant either. A reader comparing Alibaba’s 72.5 with Anthropic’s 51.4 would be comparing numbers that may not be on the same scale.
Fallbacks are part of what is measured. Artificial Analysis labels its Fable 5.2 configuration “Max, Default Fallback,” the model as served with Anthropic’s safeguard fallback active. Agents’ Last Exam instead excludes Fable’s nine downgraded tasks. One measures what a customer receives by default, the other what the model does when not blocked.
Independent results: Agents’ Last Exam
Agents’ Last Exam v1.1 is a 122-task benchmark of long-horizon professional work run inside named agent harnesses. Its public leaderboard, updated on 16 November 2026 and checked on 18 November, did not yet include GPT-6.2 Astra. Its results for the other two models, and for GPT-6 Astra as the nearest OpenAI reference, are below.
| Leaderboard entry | Harness | Effort | Pass rate | Score | Estimated cost |
|---|---|---|---|---|---|
| Claude Fable 5.2 | Claude Code | Max | 46.9% | 72.8% | $6,785.69 |
| Qwen4-Max | Codex | XHigh | 46.7% | 74.4% | $1,201.45 |
| GPT-6 Astra | Codex | Not shown | 42.6% | 73.1% | $1,097.06 |
The leaderboard says Fable 5.2 “excludes 9 tasks that are downgraded,” so its averages cover 113 tasks rather than 122. Its estimated cost is the sum of average API cost per task using recorded token usage and standard rates as of 16 November, with cache reads and writes priced separately. Dividing Fable’s $6,785.69 by 113 tasks gives $60.05 per task, which matches the figure OpenAI plotted for Fable in its launch chart. GPT-6.2 Astra’s $2.89 per task in that chart is 4.8% of $60.05, which is the basis for OpenAI’s “5% of the cost” line. Both calculations are ours.
On this benchmark Qwen4-Max looks much stronger than on the Artificial Analysis index. Its pass rate is within 0.2 points of Fable 5.2’s and its partial-credit score is higher, at about 18% of Fable’s estimated cost. It was run in OpenAI’s Codex harness rather than Alibaba’s own tooling. The leaderboard’s definitions explain why rankings on pass rate and score can differ: a model can complete more of each task while finishing fewer tasks perfectly.
With 113 to 122 tasks, one task is close to a percentage point, so a two-tenths gap is a tie. Our analysis of benchmark sizes and statistical noise explains why.
Speed and token efficiency
All three labs market efficiency, mostly against their own predecessors. OpenAI says GPT-6.2 Astra used “22% fewer output tokens” than GPT-6 Astra on GDPval v2, its internal knowledge-work benchmark; its chart there plots GPT-6.2 Astra at 50.4% with 35,812 output tokens per task and Fable 5.2 at 44.8% with 222,344. Anthropic says Fable 5.2 generates outputs “up to 30% faster” than Fable 5.1 and, at high effort on FrontierCode 2.0, “scored about level with Opus 5.5 at Max effort, for roughly half the cost per task.” Alibaba says Qwen4-Max uses about 24% and 35% fewer tokens than Qwen3.8-Max on Work360 at Xhigh and High effort, and about 48% and 49% fewer on Terminal-Bench 4.0.
The Artificial Analysis token counts are the only independent measure covering all three. GPT-6.2 Astra was the most economical in output tokens, Qwen4-Max used about twice as many and Fable 5.2 at maximum effort about four times as many. Fable generated tokens fastest, at 112 per second, and still took longest per task, at 40.8 minutes. These are maximum-effort figures. OpenAI’s own charts show lower effort settings costing much less for modest losses in score, for both its model and Fable.
What remains unknown
- GPT-6.2 Astra on Agents’ Last Exam. OpenAI’s 78.3% rests on its own run; the public leaderboard had not added the model, so its harness and cost accounting cannot yet be checked.
- Head-to-head vendor evaluations. Alibaba did not compare Qwen4-Max with either rival, Anthropic’s table predates both, and no vendor chart includes all three.
- Qwen4-Max’s regional prices. Model Studio’s layout does not tie each price unambiguously to a region, and Alibaba says pricing varies by region.
- Open-weight Qwen4 models and Daybreak access. Alibaba gives no date, size or licence for the smaller models, and OpenAI gives no date for GPT-6.2 on Daybreak.
- Fable’s cached share. Artificial Analysis does not split Fable’s 28.3 billion input tokens into cache hits and misses, so its $26.54 per task cannot be recomputed.
A leaderboard run of GPT-6.2 Astra on Agents’ Last Exam, Artificial Analysis results at lower effort with cache hits reported, or any evaluation of all three models on the same tasks and harness would change the picture.
How to read the comparison
On price alone, the ranking is simple. Qwen4-Max is far cheaper per token than the other two. Fable 5.2 and GPT-6.2 Astra share list prices, Fable has the cheaper cache reads and flat long-context pricing, and GPT-6.2 Astra has the longer default cache lifetime.
On cost per task, the ranking changes. In the only independent run covering all three, GPT-6.2 Astra and Qwen4-Max cost roughly the same per task, and Fable 5.2 at maximum effort cost more than six times as much. A team choosing between the two US models on price should measure tokens per task on its own workload at the effort level it will actually use, not compare price cards.
On capability, the independent evidence does not produce one winner. Fable 5.2 leads the broad Artificial Analysis index and its knowledge-work evaluation. GPT-6.2 Astra leads the scientific-terminal benchmark that OpenAI emphasised. Qwen4-Max trails both on the Artificial Analysis index, yet matches Fable’s pass rate on Agents’ Last Exam at a fraction of the estimated cost. The benchmark each lab chose to lead with is, in each case, one on which its model does well.
Sources
- OpenAI: Introducing GPT-6.2 Astra and Sol (17 November 2026)
- OpenAI API docs: GPT-6.2 Astra model page
- OpenAI API docs: GPT-6 Astra model page
- OpenAI API docs: Using GPT-6.2
- OpenAI: GPT-6.2 Astra and Sol System Card (17 November 2026)
- Qwen Team: Qwen4-Max, Your New Teammate Has Arrived
- Alibaba Cloud: Alibaba Unveils Qwen4-Max (16 November 2026)
- Alibaba Cloud: Alibaba Cloud Brings Qwen4-Max to the Enterprise (17 November 2026)
- Alibaba Cloud Model Studio: Qwen4-Max model page
- Alibaba Cloud Model Studio: Model pricing
- Alibaba Cloud Model Studio: Explicit cache for Qwen4
- QwenCloud: Qwen4-Max model page
- Anthropic: Introducing Claude Fable 5.2 and Claude Mythos 5.2 (9 November 2026)
- Claude Platform docs: Models overview
- Claude Platform docs: Pricing
- Artificial Analysis: LLM leaderboard
- Artificial Analysis: GPT-6.2 Astra
- Artificial Analysis: Qwen4 Max
- Artificial Analysis: Claude Fable 5.2
- Artificial Analysis: Terminal-Bench-Science 0.1
- Agents' Last Exam v1.1 leaderboard
Questions readers ask
Which of the three models is cheapest per token?
Qwen4-Max. Alibaba Cloud lists it at $1.50 per million input tokens and $4.50 per million output tokens on its Singapore (international) deployment, with cache reads at $0.075. GPT-6.2 Astra and Claude Fable 5.2 both list $8 input and $40 output, with cached input at $0.40 and $0.20 respectively.
Do identical list prices mean GPT-6.2 Astra and Fable 5.2 cost the same to run?
No. Cost per task depends on how many tokens a model uses. On the Artificial Analysis Intelligence Index, Fable 5.2 at maximum effort averaged $26.54 per task against $4.00 for GPT-6.2 Astra at maximum effort, mainly because Fable generated about four times as many output tokens.
Are any of these models open-weight?
No. All three are proprietary and served through their vendors' APIs and cloud partners. Alibaba says smaller, open-weight Qwen4 models are coming, and it open-sourced the Qwen-Live-Delivery-Plugin, but Qwen4-Max's weights are not published.
Which model scores highest on independent tests?
It depends on the test. Artificial Analysis puts Fable 5.2 first on its Intelligence Index (65) ahead of GPT-6.2 Astra (62) and Qwen4-Max (43), while GPT-6.2 Astra leads its Terminal-Bench-Science run. On the Agents' Last Exam leaderboard, Fable 5.2 and Qwen4-Max are nearly tied on pass rate; GPT-6.2 Astra was not yet listed when The Frontier Wire checked.
Did The Frontier Wire test these models?
No. This comparison reads the vendors' launch posts, model and pricing documentation, Artificial Analysis's published results and the Agents' Last Exam leaderboard as retrieved on 18 November 2026. Cost examples are our arithmetic on list prices.
