Qwen4-Max or Claude Fable 5.2 for your work?
Start with the job you need finished
These models pose a useful trade-off for a team buying AI work. Qwen4-Max has a lower standard API token price. Fable 5.2 has stronger results across much of the shared evaluation data. The question is whether the extra capability changes the quality of a finished job enough to justify its cost.
In the current matching tracks, Qwen leads on 0, Fable leads on 7, and 0 are tied, across 7 comparisons. This is a count of reported benchmark outcomes, not a blended capability score. Some tracks cover related skills, so they are not independent votes in an overall election.
When a lower-cost attempt is easy to check
Try it on bounded coding tasks, first drafts, structured extraction and repeatable production steps where you can quickly tell whether the output works. Its lower rates give you room to run a trial before committing the rest of a workflow.
When the hard part determines the value
Prioritize it for demanding reasoning, complex repository changes and polished deliverables where a failed attempt or an hour of correction is expensive. Evaluate the effort setting as well as the model.
A practical setup may use both: a lower-cost model for well-defined steps and a stronger model for difficult reasoning or final revision. Judge that setup by the total time and money needed for an acceptable result.
Every shared benchmark, explained
The following 7 tracks come from the same data and matching rules as the basic comparison. We keep benchmark version, tool access and evaluation cohort aligned. Within the latest reported date, that page selects the strongest reported score for each model. Consequently, a model can appear at different effort settings across the charts. Read the labels and source notes before drawing a cost conclusion.
Business deliverables: quality matters after the answer is written
AA-Briefcase combines analytical and presentation quality into an Elo rating. It covers complex knowledge work with source files and linked deliverables. Elo is a relative rating, not a percentage of projects completed.
For a spreadsheet or deck that a client will use, inspect both correctness and usability. A stronger model may earn its higher token price if its output needs less reconstruction before handoff.
Higher is better. Elo ratings are relative, not percentage success rates.
Work automations: score the whole workflow yourself
AutomationBench-AA reports objectives completed without guardrail violations on a private 657-task set. It does not use the strict whole-task completion metric from the separate Zapier track.
Test the complete sequence: finding a record, making the right decision and preparing or executing the next step. Count wrong actions and unfinished handoffs, not just correctly completed intermediate objectives.
Higher is better. Percentage scores apply to this evaluation, not arbitrary work.
Mergeable code: look beyond a plausible patch
FrontierCode 2.0 in this comparison uses the independent mini-swe-agent protocol. The target is useful repository work; it is not a direct measure of how many hours an engineer saves.
For a difficult feature or bug, evaluate whether the change is mergeable after tests and review. Try the cheaper model on bounded changes first; use the higher-performing option where failed attempts consume valuable engineering time.
Higher is better. Percentage scores apply to this evaluation, not arbitrary work.
Document reasoning: small differences deserve a real-file trial
GDP.pdf asks models to reason over professional documents and uses an all-pass metric across five attempts. This is stricter than getting some questions or rubric items right.
Use a representative folder of your own PDFs and require citations to the source material. A narrow score difference is a reason to compare errors and review time, rather than declare a universal winner. Effort settings can change the result.
Higher is better. Percentage scores apply to this evaluation, not arbitrary work.
Hard reasoning: reserve it for the difficult part of the job
HLE-Diamond tests difficult multidisciplinary questions without tools. It is useful evidence about reasoning under this protocol, rather than how well a model will browse, use your CRM or finish a client deliverable.
If your workflow repeatedly stalls on a hard analytical step, give the stronger reasoning model that step and measure whether the improvement survives on your own questions. Routine extraction can stay on a cheaper model.
Higher is better. Percentage scores apply to this evaluation, not arbitrary work.
Terminal coding: test the agent and its environment together
This track uses Terminal-Bench 4.0 with mini-swe-agent and three repeats. It is separate from providers' native coding-agent launch results.
Use the same repository, tests, tool permissions and task brief for both models. Count successful fixes, regressions and human interventions. A model's token price alone does not tell you the cost of a working change.
Higher is better. Percentage scores apply to this evaluation, not arbitrary work.
Scientific workflows: keep the experimental checks
Terminal-Bench-Science evaluates scientific work in a terminal. These independent runs have their own harness and repeat policy and should not be merged with native-agent system-card results.
For computational research, require the code, parameters and outputs needed to reproduce the result. Use benchmark performance to choose what to test, then verify the actual experiment before relying on its conclusion.
Higher is better. Percentage scores apply to this evaluation, not arbitrary work.
All shared benchmark scores and sources
| Benchmark and setup | Qwen4-Max | Claude Fable 5.2 |
|---|---|---|
| AA-Briefcase1.1 / combined Elo · Tools enabled | 1,686.07 Eloxhigh | 1,785.32 Elomax |
| AutomationBench-AA1.0.6 / AA private 657 · Tools enabled | 72.6%xhigh | 76.4%max |
| FrontierCode2.0 / independent mini-swe-agent · Tools enabled | 26.5%xhigh | 49.1%high |
| GDP.pdf100 tasks / All-pass / five attempts · No tools | 26.2%xhigh | 26.8%low |
| HLE-Diamond1,000 questions / no tools · No tools | 44.9%xhigh | 80.2%max |
| Terminal-Bench4.0 / mini-swe-agent · Tools enabled | 48.5%xhigh | 61.6%xhigh |
| Terminal-Bench-Science0.1 / independent mini-swe-agent · Tools enabled | 23.3%xhigh | 73.3%max |
Token price is the starting point
At the standard first-party offerings selected in the basic comparison, Qwen’s input/output rates are $1.50 / $4.50 per million tokens. Fable’s are $8.00 / $40.00. Cached input, optional fast modes and other hosting providers can change a specific bill.
Blended token cost
USD per 1M tokens · 50% input and 50% output · lower is better
A concrete workload
Take a run that reads 100,000 uncached input tokens and generates 20,000 output tokens. At these rates, its token cost is $0.24 on Qwen and $1.60 on Fable. Across 1,000 identical token workloads, that becomes $240.00 versus $1,600.00. This illustration holds token use constant; real models can use different amounts of reasoning and output.
| Model | One attempt | 1,000 attempts |
|---|---|---|
| Qwen4-Max | $0.24 | $240.00 |
| Claude Fable 5.2 | $1.60 | $1,600.00 |
For a business, add tool charges, retries and review time. A cheap answer that needs to be rebuilt can cost more than an expensive answer that is ready to use. Track cost per accepted deliverable or merged change, rather than choosing from a token-price chart alone.
Context, output and provider details
Both catalog entries list a one-million-token context window. This is a capacity limit, not a promise of perfect recall. Large repositories and document collections still need a clear brief, relevant files and checks that the model used the right evidence.
| Specification | Qwen4-Max | Claude Fable 5.2 |
|---|---|---|
| Provider selected | QwenCloud | Anthropic |
| Context size | 1,000,000 tokens | 1,000,000 tokens |
| Knowledge cutoff | Not reported | 2026-06 |
| Input price / 1M tokens | $1.50 | $8.00 |
| Output price / 1M tokens | $4.50 | $40.00 |
| Exact maximum output | 262,144 | 128,000 |
| Selected endpoint TTFT | Not reported | Not reported |
| Selected endpoint output speed | Not reported | 103.417 tokens/s |
All available provider offerings
| Model / provider | Configuration | Input / 1M | Output / 1M | Context | TTFT, seconds | Tokens/s |
|---|---|---|---|---|---|---|
| Qwen4-MaxQwenCloud | Standard API, Singapore; uncached input | $1.50 | $4.50 | 1,000,000 | Not reported | Not reported |
| Claude Fable 5.2Anthropic | Fast mode API | $16.00 | $80.00 | 1,000,000 | Not reported | Not reported |
| Claude Fable 5.2Anthropic | Standard uncached API | $8.00 | $40.00 | 1,000,000 | Not reported | 103.417 |
Alibaba’s released Qwen4-Max API documentation lists 131,072 output tokens without thinking and 262,144 combined thinking and final-answer tokens. The table records the combined budget; thinking consumes some of that room. These exact API limits conflict with the marketplace’s rounded 393K label, which we do not use as the output ceiling. The separate Preview model has its own limits. Fable’s recorded standard API output limit is 128,000 tokens. Qwen’s knowledge cutoff remains undisclosed in the reviewed official sources; Fable’s catalog entry lists June 2026.
The default provider is part of the comparison. Qwen is shown through QwenCloud and Fable through Anthropic. The interactive page lets you change the provider and inspect its price and endpoint measurements. Those choices can differ from the least expensive offering shown elsewhere on the site.
Speed and latency depend on effort
The first-party workload snapshot contains five Fable effort levels and no matching effort-level measurements for Qwen4-Max. The charts therefore show Qwen as unreported. We also checked Artificial Analysis’s Qwen4-Max provider records on November 16: output speed, time to first token and time to first answer are all unreported there. A missing measurement is not evidence that it is slower.
Time to first answer includes thinking. It is distinct from the endpoint time-to-first-token field in the specification table. An agent’s total job duration also includes tool calls, test execution and revisions, so neither number alone predicts when a feature or presentation will be finished.
Effort-level measurements and sources
| Model | Effort | Tokens/s | First answer, seconds | Source |
|---|---|---|---|---|
| Qwen4-Max | No matching effort-level measurements | |||
| Claude Fable 5.2 | low | 92.416 | 8.731 | Measurement ↗ |
| Claude Fable 5.2 | medium | 88.361 | 13.445 | Measurement ↗ |
| Claude Fable 5.2 | high | 93.001 | 30.793 | Measurement ↗ |
| Claude Fable 5.2 | xhigh | 103.4 | 73.442 | Measurement ↗ |
| Claude Fable 5.2 | max | 102.829 | 235.061 | Measurement ↗ |
The Agent Index is incomplete for Qwen
Fable has 7/7 required components and a score of 89.594. Qwen currently has 4/7. Its missing LiveBench Agentic, Coding and Browser Use results prevent a valid composite score. We do not fill those gaps with its DeepSWE, Work360 or Terminal-Bench results because they measure different things.
Agent Index coverage and score
Seven required benchmark components · missing scores are not zero
- Qwen4-Max: Needs all 7 results (4/7 available).
Use the shared individual benchmarks to make a narrower comparison today. The existing updater is registered for Qwen4-Max’s LiveBench model ID and can import matching scores when they are published.
A small trial will tell you more than another leaderboard
Pick a handful of jobs you already understand: a bug with a known reproduction, a workbook with a checked answer, and a document task with a clear rubric. Give both models the same source material and tool access. Record the effort setting, token spend, elapsed time and the edits required before the result is usable.
For fixed-fee client work, review time can be the deciding cost. For frequent internal automations, the cost of repeated attempts may matter more. For a personal project, a cheaper model that gets you to a working first version may be enough. The best choice is the model and configuration that make the complete job worthwhile.
If Qwen gets the routine steps right and Fable resolves the difficult failures, keep both in the workflow. If one model consistently reduces total rework, the token bill is only one part of that result.
One factual comparison, two useful views
This report and the quick comparison read the same model catalog, provider offerings, benchmark cohorts and performance snapshot. The quick page is for checking facts and changing providers. This report is for understanding the trade-offs and designing a trial.
The shared charts match benchmark versions and evaluation cohorts. Their selection policy uses the latest reported date, then the highest score within that date. Effort settings are retained beside each result, and no overall capability score is calculated from the win count. The Qwen release guide uses a separately labelled fixed-effort peer view; its numbers can differ where that view selects Max rather than another effort level.
Download the complete report data and original source links as JSON ↗. Every benchmark has source links and HTML data tables, so readers and agents can inspect the data without operating a chart.
