
Qwen4-Max vs Claude Fable 5.1: Six Index Points for an Eleventh of the Output Price
- 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 · 96 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-0717 tok/s
- OrcaNEWOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 87 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 · 72 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 130 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 65 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 92 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 · 193 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 150 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 · 120 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 · 219 tok/s
Qwen4-Max is Alibaba's new flagship and it went generally available on 16 November 2026 at $1.50 per million input tokens and $4.50 per million output. Claude Fable 5.1 has been on sale since 1 September 2026 at $10.00 and $50.00. On the token card that is a 6.7× gap on input and 11.1× on output, and the question the gap raises is what six points of intelligence are worth: 43 for Qwen4-Max against 49 for Claude Fable 5.1, both read on the same revision of the Artificial Analysis Intelligence Index on the day of the Qwen launch. The answer that survives contact with an invoice is smaller than the token ratio suggests — $3.27 per completed index task against $12.99 — because the cheaper model writes almost twice as many tokens to arrive at the same answer.
The task price, which is not the token price
Artificial Analysis prices a model two ways and only one of them is a rate card. The first is per million tokens. The second is what the model actually spent getting through the evaluation, which folds the rate card into how much the model chooses to write. Those two orderings disagree in this matchup, and the disagreement is the comparison.
• Token price, input — Qwen4-Max $1.50 per million against Claude Fable 5.1 at $10.00. Ratio 6.7×.
• Token price, output — $4.50 against $50.00. Ratio 11.1×.
• Cost per Intelligence Index task — $3.27 against $12.99. Ratio 4.0×, not 11.1×.
• Output tokens generated during the index run — 127 million for Qwen4-Max, which the tracker ranks 65th of 80 on its verbosity scale and calls very verbose against a median of 78.6 million; 9.9 billion input tokens went in.
• Intelligence Index — 43 for Qwen4-Max, ranked 38th of 81 models on the board; 49 for Claude Fable 5.1, ranked 26th. Same index revision, read on 16 November 2026.
The mechanism is unglamorous. A reasoning model that thinks longer and restates more of its working spends part of its price advantage on tokens it did not need to emit. Qwen4-Max is still far cheaper per finished task — a quarter of Claude Fable 5.1's bill on this measurement — but it is cheaper by four times, not by eleven, and any comparison quoting only the per-token card is quoting the wrong number. Where output length is pinned by a schema, a strict answer format or a hard max-token cap, the verbosity penalty shrinks and the token ratio reasserts itself. Where the workload is open-ended analysis, expect something close to the four.

What the two cards bill, line by line
Every figure below is the vendor's own list price as published on 16 November 2026. Neither vendor's card has been audited by anyone, and Alibaba's rates are quoted in US dollars here because that is how the Singapore and Global cards are presented.
• Input — Qwen4-Max $1.50 per million in Singapore, $1.413 in the other five regions; Claude Fable 5.1 $10.00.
• Output — Qwen4-Max $4.50 in Singapore, $4.24 elsewhere; Claude Fable 5.1 $50.00.
• Cache read — Qwen4-Max $0.075 per million, which is 5% of its input rate; Claude Fable 5.1 $0.25 per million, which is 2.5% of its input rate. The independent tracker records these as 95% and 98% cache discounts, and it is Fable that wins that line proportionally.
• Cache write — Qwen4-Max $1.875 per million for a five-minute window and $3.00 for an hour; Claude Fable 5.1 $12.50 and $20.00. Both charge a premium to write, so a prefix that changes every turn costs money on either model.
• Batch — Claude Fable 5.1 halves both directions to $5.00 and $25.00. Qwen4-Max offers a batch path in Beijing only, where Batch File bills $0.707 and $2.12 and Batch Chat bills the real-time rate; the 50% discount attaches to Batch File, not to the word "batch" in the endpoint name.
• Context and output — both serve 1,000,000 tokens. Claude Fable 5.1 caps output at 128,000. Qwen4-Max caps non-thinking output at 131,072 and puts thinking and the final answer inside one 262,144-token budget, so its effective answer ceiling in thinking mode is whatever the reasoning left behind.
• Input — Qwen4-Max accepts text, images and video. Claude Fable 5.1 accepts text and images. That is the one row where the cheaper model does something the expensive one cannot.
One structural difference hides in the cache line. Fable's cache read is proportionally cheaper, so a workload built on a large fixed prefix — a system prompt, a tool schema, a document set re-read every turn — narrows the cost gap further than the headline suggests. A workload with no cache reuse is priced exactly at the token ratios above.
Where the 262,144-token budget actually bites
Qwen4-Max ships with reasoning on and effort set to high. Omitting the parameter is not a neutral act: reasoning_effort accepts low, high and xhigh and defaults to high, thinking_budget defaults to 131,072, and max_completion_tokens defaults to 262,144 in thinking mode, which is the entire shared budget. There is no middle setting either — the Chat Completions and DashScope paths map a request for medium onto high, so asking for medium costs what high costs.
Claude Fable 5.1 uses adaptive thinking that cannot be turned off, with an effort dial from low to max and a default of high. Its 128,000 output tokens are a clean ceiling rather than a shared bucket, which matters when you are budgeting an agent loop: on Qwen4-Max the reasoning tokens and the answer compete for the same 262,144, and a long tool-calling run can spend the budget before it writes anything you can use.
There is also a migration trap specific to this model. The preview identifier qwen4-max-preview carried a 393,216-token shared budget and is documented for retirement at 24:00 on 25 November 2026 UTC+8. Anything tuned against the preview now has a third less headroom and nine days to notice, and neither change raises an error when it is wrong.
Claude Fable 5.1 is no longer the current Fable
This matters more than any row above. Anthropic shipped Claude Fable 5.2 on 9 November 2026, a week before Qwen4-Max, at $8.00 and $40.00 — a 20% cut on the 5.1 card — and it tops the same index at 65. So there are two honest framings of this matchup and they point differently.
• Against Claude Fable 5.1 — the model named in this comparison and the one still on many contracts — Qwen4-Max is 6.7× cheaper on input, 11.1× cheaper on output, 4× cheaper per index task, and 6 points behind on capability.
• Against Claude Fable 5.2 — the current Fable — Qwen4-Max is 5.3× cheaper on input and 8.9× cheaper on output, and 22 points behind on capability. The cheap-token argument survives; the "close enough on the index" argument does not.
If you are choosing a default for a new workload, the model to price against is 5.2, and the Qwen4-Max case becomes a case about price per unit of capability rather than about near-parity. If you are choosing whether to keep paying for 5.1, the answer is separate and probably not about Qwen at all.
Refusals, regions, and the operational surface
Claude Fable 5.1 runs with production safeguards that can decline a request outright, and Anthropic's own guidance is that a refusal comes back as a normal HTTP 200 whose stop reason is refusal rather than as an error — code that only catches exceptions will silently treat a refusal as a completed call. Qwen4-Max has no comparable published behaviour; what it does publish is a regional map, with per-token metering in Beijing, Singapore, Frankfurt, US Virginia, Tokyo and Hong Kong, and batch inference restricted to Beijing.
The protocol surface is broader on Qwen4-Max than its provenance suggests: OpenAI Chat Completions, OpenAI Responses, Anthropic Messages and DashScope, all four, with structured output, parallel tool calling, prefix completion, web search and both explicit and implicit caching. A team already speaking Anthropic's Messages format can move a request body across without rewriting it, which is not something most Alibaba launches of the last two years could claim.
Calling both from one key
Claude Fable 5.1 is routable on OrcaRouter at Anthropic's list price with 0% markup — $10.00 in and $50.00 out per million tokens, with the cache rates above — and Claude Fable 5.2 sits on the same endpoint at $8.00 and $40.00. One API for 200-plus models means the version question in the section above is a routing change rather than a migration: the same request body can be pointed at 5.1 for a pinned production path and at 5.2 for a slice, and a provider price change lands here the day the vendor publishes it.
Qwen4-Max is not in the OrcaRouter catalogue as of 16 November 2026. It is reachable through Alibaba's own Model Studio API and several third-party platforms, which means the candidate arm of this comparison is a separate contract whichever way you reach it. That asymmetry is the practical story of the matchup: your incumbent is already one key away, the challenger is not, and the sensible move is to keep the incumbent serving volume while a shadow slice of traffic — priced per completed task, not per token — decides whether a 4× bill reduction is worth six index points, or twenty-two.

What would change this verdict
One column on the independent page is empty. Artificial Analysis has no performance reading for Qwen4-Max yet: no output speed, no time to first token, no minutes per index task. Claude Fable 5.1 has all three, and they are not flattering — 147.58 seconds to first answer token, 89 tokens per second, 22.1 minutes per task — but they exist, and the absence on the Qwen side means the 4× cost advantage is currently a claim about money per completed task with no corresponding claim about time. A model that is four times cheaper and twice as slow is not a four-times saving on anything interactive.

Two other things to watch, both dated. Alibaba's own claim that average cost per task fell from $4.43 to $1.67 — a 62% reduction — is vendor arithmetic that multiplies a price cut by a token-efficiency claim, and it is the one number in this piece that no outside party has reproduced. And the 25 November preview shutdown will produce a wave of migrations onto the GA identifier with a smaller output budget than the one they tuned against; if you are in that wave, the six points and the four times both matter less than noticing the 262,144 before a long run does.
Compared in this article2
Detected from this article · Benchmarks: Artificial Analysis · updated daily
