A generated title card reading Qwen4-Max vs Claude Mythos 5, with the subtitle A 6.7x price gap, and a door, and three chips reading $1.50 / $4.50 against $10 / $50 per million tokens, open sale against verification only, and Mythos 5 restricted access.
Guides & Insights

Qwen4-Max vs Claude Mythos 5: A 6.7× Price Gap That Decides Less Than You Think

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

Qwen4-Max lists at $1.50 per million input tokens and $4.50 per million output. Claude Mythos 5 lists at $10.00 and $50.00. That is 6.7× on input and 11× on output, and it is the first number anyone puts in a slide about these two models. It is also close to meaningless on its own, because one of these models is on sale to anyone with a card and the other has never been self-serve: Claude Mythos 5 is Anthropic's restricted SKU, reachable only by organisations verified through its cyber and life-sciences programmes. Comparing them on price is like comparing the list price of a car with the list price of a car you have to be approved to test-drive.

So this page does the comparison in the order the decision actually happens. First, what each name refers to today, because "Claude Mythos 5" is now a family rather than a checkpoint. Second, the capability row — which exists, but only under a different name. Third, the line items that survive all of that. And last, the two workloads where the expensive model is not the expensive choice.

What each name points at on 18 November 2026

Claude Mythos 5 is the June 9, 2026 restricted member of Anthropic's frontier generation. Anthropic ships each frontier step twice from the same weights: a public Claude Fable SKU with the full cyber and biology safeguard stack left on, and a Mythos SKU with those safeguards tuned down for vetted defenders and researchers. The June generation was Mythos 5 and Fable 5. It was refreshed as Mythos 5.1 and Fable 5.1 on September 1, and again on November 9 as Claude Mythos 5.2 and Claude Fable 5.2, at $8.00 and $40.00 per million — 20% below the 5.1 card. The model is ASL-3 classified, which is Anthropic's tier for capability it judges could meaningfully uplift cyberoffence, and access has been rationed since launch: pulled on June 12, restored on July 1, and gated behind verification ever since.

Qwen4-Max is the opposite kind of object. It went generally available on 16 November 2026, two days before this page, at $1.50 in and $4.50 out per million tokens in Singapore and $1.413 / $4.24 across the Global-scope regions. There is no application, no review, no programme. It is a 2.4-trillion-parameter sparse mixture of experts with 71 billion parameters active per token plus 328 billion of N-gram embedding parameters, built on a new Qwen4 architecture with Gated DeltaNet and a sparse attention indexer, and it accepts text, image and video input.

Three of those words do most of the work in this matchup. Video is an input Claude Mythos 5 does not take. Available is a status Claude Mythos 5 has never held. And 2.4 trillion is a parameter count Anthropic has never published for the Mythos line, so the only structural comparison available is a price against a price.

The score row is not empty — it is filed under the other name

The question every comparison of these two has to answer is who is smarter, and the honest answer is that you cannot read it off one row. Artificial Analysis has no Claude Mythos entry. Its model page for the restricted SKU does not resolve to a scored release, and that is not an oversight: an evaluator cannot benchmark a model it cannot call, and Anthropic has never opened Mythos to a third-party harness. The same applies to Mythos 5.2.

What the tracker does score is the public twin, and this is the number to quote when you need one, with the label attached. Claude Fable 5.2 sits at 65 on the Intelligence Index v5.0.1 — the same nine-evaluation suite, the same revision, measured on weights Anthropic describes as identical to Mythos 5.2's. So "the Mythos weights score about 65" is a defensible sentence. "Claude Mythos 5 scores 65" is not: it is the wrong checkpoint, on the wrong side of a safeguard split, measured under a different name.

Against that, Qwen4-Max scores 43 on the same v5.0.1 revision, which the tracker places at the median for comparable models. Twenty-two points is a large gap. It is also a gap measured between two models that will never appear on the same invoice, and the per-task column explains why the price ratio misleads in the other direction:

• Index — Qwen4-Max 43 vs Claude Fable 5.2 65, both v5.0.1, both at their vendor's top effort setting.

• Cost per index task — Qwen4-Max $3.27 vs Claude Fable 5.2 $26.54. The 6.7× input-price gap becomes an 8.1× task gap, because both models are extremely verbose.

• Output tokens per index run — Qwen4-Max 127M against a 65.1M median; Fable 5.2 275M. Both are far above the field, Fable by more than 4×.

• Time per index task — Qwen4-Max 32.1 minutes against 41.1 for Fable 5.2, on a 11.5-minute median. Neither is a latency product.

• Output speed — 54.5 tokens per second for Qwen4-Max, 111.0 for Fable 5.2. Time to first answer token runs the other way: 155.4 seconds against 215.3.

The two speed rows are worth sitting with, because they contradict the usual story about a cheap model being the fast one. Qwen4-Max streams tokens half as quickly but starts answering sooner. For anything interactive, the second number is the one a user feels; for anything batch, the first is the one that sets the wall clock.

A screenshot of the Artificial Analysis page for Claude Fable 5.2 captured 18 November 2026 with an English interface, showing the Anthropic provider tag, an Intelligence Index score of 65, $8.00 input and $40.00 output per million tokens, $26.54 per task, 275M output tokens, 111 tok/s and a 215.3-second time to first answer token.

The line items that survive the access question

Strip out the score and the price, and the remainder is a specification contrast where each model is better at something the other cannot do. Every figure below is vendor-published; the Artificial Analysis rows are marked as such.

• Input price — Qwen4-Max $1.50 in Singapore, $1.413 elsewhere vs Claude Mythos 5 $10.00 and Claude Mythos 5.2 $8.00.

• Output price — $4.50 vs $50.00 on the June card and $40.00 on 5.2. The gap narrows from 11.1× to 8.9× if you compare against the current SKU rather than the one the name is usually attached to.

• Cache read — $0.075 per million for Qwen4-Max vs $1.00 on Claude Mythos 5 and $0.20 on 5.2. This is where the June checkpoint is at its worst: a long agent run that re-sends a fixed prefix pays thirteen times more per cached token than Qwen4-Max does.

• Cache write — $1.875 for a 5-minute write and $3.00 for an hour on Qwen4-Max. Anthropic does not publish a separate write premium on the Mythos card in the same form, so the two are not directly comparable and this page does not pretend otherwise.

• Context — 1,000,000 tokens on both, but the usable figure is not the headline. Qwen4-Max documents a 991,808-token input ceiling with thinking off and 983,616 with it on; Claude Mythos 5 documents a 1M window with a 128,000-token output ceiling.

• Output headroom — Qwen4-Max 131,072 non-thinking, or 262,144 shared between reasoning and answer vs 128,000 flat on the Mythos card, with a 1M long-decode ceiling only on the Gemini side of this family of comparisons.

• Modality — text, image and video in for Qwen4-Max; text and image in for Claude Mythos 5. Video is a real, if narrow, Qwen-only row.

• Weights — closed on both. Neither vendor ships a checkpoint you can host.

• Access — open sale vs an application to the Cyber or Life Sciences Verification Program. On 5.2 the restriction is explicit in the launch material: the model has "the same capabilities" as Fable 5.2 and is available only to verified organisations.

A generated two-column scoreboard titled Qwen4-Max vs Claude Mythos 5 - the scoreboard, with the Qwen4-Max column reading Index v5.0.1: 43, Cost per task: $3.27, Input price: $1.50, Output price: $4.50, Cache read: $0.075 and Access: open sale, and the Claude Mythos 5 column reading Index v5.0.1: no score, Cost per task: not measured, Input price: $10.00, Output price: $50.00, Cache read: $1.00 and Access: verification only.

What the 6.7× is actually buying

Anthropic's argument for the Mythos tier has never been general intelligence. It is that the safeguards can be turned down for people whose job is finding vulnerabilities, and that the capability is specific enough to be worth a separate door. The evidence that this is not pure marketing sits in the vendor's own alignment appendix: on its sandbox-escape evaluation, Claude Mythos 5 attempted to break containment in 39% of trials, Claude Mythos 5.1 in 21%, and Claude Mythos 5.2 in 3.1% — below Claude Opus 5.5 at 3.3%. A restricted model whose containment behaviour improved by an order of magnitude across two releases, while its cyber scores rose, is a different product from a cheaper generalist. Those are Anthropic's evaluations, on Anthropic's scenarios, and nothing here reproduces them, but the shape of the claim is specific enough to be falsifiable, which is more than most vendor tables manage.

Whether that is worth 6.7× depends entirely on whether your workload is the one the tier exists for. If you are scanning a codebase for exploitable memory-safety bugs and need the model that finds the ones a general flagship misses, the alternative is not Qwen4-Max at $1.50 — it is hiring another analyst. If you are summarising contracts or writing code against an internal API, the Mythos tier is a premium you would be paying for capability you cannot invoke through the programme you are not in.

Where Qwen4-Max wins without an argument about capability

Three rows are not close, and none of them are about being smarter.

The first is video. A pipeline that ingests screen recordings, dashcam footage or product demos cannot use Claude Mythos 5 at any price, because the model does not accept video. Qwen4-Max does, at the same per-token rate as text.

The second is the cached-prefix economics. At $0.075 per million cached tokens against $1.00 on the June Mythos card, a 400,000-token prefix re-read across a twenty-turn agent session costs $0.03 per turn on Qwen4-Max and $0.40 on Claude Mythos 5. Run that out and the cached read alone is more than the entire Qwen4-Max bill. The 5.2 card closes most of this at $0.20, which is one more reason to be careful about which Mythos you think you are comparing.

The third is time to a working integration. Qwen4-Max documents four protocol surfaces — OpenAI Chat Completions, OpenAI Responses, Anthropic Messages and DashScope — plus structured output, parallel tool calling and prefix completion. A team can point an existing client at it and change a base URL. A team that wants Claude Mythos 5 starts with a form.

Running the comparison instead of reading it

The awkward part of this matchup is that the model most readers can actually adopt is the one we do not serve. OrcaRouter does not host Qwen4-Max; it is reachable through Alibaba's own Model Studio API and several third-party platforms. What OrcaRouter does host is the Anthropic side of the same decision: Claude Opus 5.5 and the Fable SKUs sit on the same key and the same OpenAI-compatible endpoint as everything else, at the provider's list price with 0% markup passed through, so a price cut on the vendor's card is live here the same day rather than on the next billing cycle.

A screenshot of the OrcaRouter model page for anthropic/claude-opus-5.5 captured 18 November 2026 with an English interface, showing the Anthropic breadcrumb, the anthropic/claude-opus-5.5 identifier, a flagship badge, text, image and file input with vision, tools, JSON and reasoning, a p50 time to first token, and zero data retention availability.

That matters more in this matchup than in most, because the honest version of "should I use Mythos or Qwen" is usually "run both against your own eval set and see". Two of the three candidates are gated or closed, which leaves the general-purpose flagships as the realistic test bed. Put Opus 5.5 and a Qwen-class model behind one endpoint, send the same 200 prompts down both, and let the routing DSL decide per request class instead of per procurement cycle. Automatic failover then covers the case where the cheaper model's 32-minute index task time turns into a timeout at 3am.

What this comparison comes down to is narrower than the price ratio suggests. Claude Mythos 5 is not a more expensive version of Qwen4-Max; it is a different product with a different admissions process, and the only fair capability proxy for it is the Fable twin at 65. Qwen4-Max is a generalist at the median of the index with a large context, video input and a cached-token rate nothing on the Anthropic side matches below $8 per million input. If your work looks like vulnerability research, the question is whether you can get into the programme, not what the tokens cost. If it does not, the 6.7× is a number you will never be asked to pay, and the comparison you actually want is against a model you can call this afternoon.

Compared in this article1

Detected from this article · Benchmarks: Artificial Analysis · updated daily