Fast-moving cluster — prices and models change

Qwen 3.8 Max is announced but not released, and every vendor here updates pricing and weights on its own schedule. Figures below are dated as of August 2026 and sourced from vendor pages; check the alternatives page for the current picture.

See the up-to-date alternatives page

Model comparison · Chinese cluster

Fable 5 vs Chinese Models

Claude Fable 5 against Qwen 3.8 Max, Kimi K3, GLM 5.2, DeepSeek V4 Pro and DeepSeek V4 Flash — price, coding benchmarks, open weights and the export-control reality, as of August 2026.

The short version

For most mainstream work Claude Fable 5 stays the reference, but the Chinese cluster competes hard on price and open weights. Fable 5 wins on agentic coding and on safety/retention guarantees; DeepSeek, GLM and Qwen win on cost and self-hosting; Kimi K3 is the closest head-to-head. Export-control timing is a real differentiator — see below.

Master matrix

Price per million tokens (input / output), context window, open-source status and best-fit, as of August 2026. Fable 5 prices come from Anthropic's published rates; competitor rates are vendor figures labeled by source.

ModelInput / $1MOutput / $1MContextOpen-sourceBest forSource
Claude Fable 5$10.00$50.001MClosed APIAgentic coding + safetyAnthropic-published
Qwen 3.8 Max≈$2.00≈$6.001MAnnounced, not releasedOpen weights (announced)Vendor-published
Kimi K3$3.00$15.001MOpen weightsClosest head-to-head qualityVendor-published
GLM 5.2~$1.40~$4.401MOpen weights (MIT)Open weights, strong codingVendor-published
DeepSeek V4 Pro$0.435$0.871MOpen weightsCheapest open agenticVendor-published
DeepSeek V4 Flash$0.14$0.281MOpen weights (MIT)Lowest-cost high volumeVendor-published

Prices are USD per million tokens, as of August 2026. Fable 5 rates are Anthropic-published; competitor rates are as published by each vendor (labeled 'vendor') — Qwen 3.8 Max rates are announced target pricing, not confirmed on a released model.

Coding benchmarks

Published numbers only. Every cell is labeled by source — 'vendor' means the figure is vendor-reported and not independently confirmed; 'not published' means the vendor has not released a number.

ModelClaude Fable 5Qwen 3.8 MaxKimi K3GLM 5.2DeepSeek V4 ProDeepSeek V4 Flash
SWE-bench Pro80.3%table.val_vendortable.val_vendor62.1%55.4%table.val_not_published
FrontierCode29.3%table.val_nonetable.val_nonetable.val_nonetable.val_nonetable.val_none

Anthropic-published: Fable 5's 80.3% on SWE-bench Pro and 29.3% on FrontierCode. GLM 5.2 (62.1% SWE-bench Pro) and DeepSeek V4 Pro (55.4%) are vendor-reported and not independently confirmed. Qwen and Kimi have not published comparable figures — labeled 'not confirmed' rather than guessed.

Benchmarks are not directly comparable across vendors — each runs its own harness and prompt set. Treat cross-model deltas as directional, not exact.

Which should you pick?

Five decision rows, each mapped to the model that fits.

  • You want self-hosting or open weights

    GLM 5.2 (MIT) and the DeepSeek V4 line are the strong picks — downloadable weights you can run on your own hardware. Qwen 3.8 Max promises the same but is not released yet.

  • You're cost-sensitive at high volume

    DeepSeek V4 Flash (~$0.14/$0.28) and V4 Pro ($0.435/$0.87) are the cheapest by a wide margin. Fable 5 is $10/$50 — justifiable when quality and guarantees matter, not for bulk throughput.

  • You want the closest quality challenger

    Kimi K3 is the closest head-to-head to Fable 5 on general coding quality, at ~$3/$15. It trails on the hardest agentic tasks, but the gap narrows on everyday work.

  • You need agentic coding and safety

    Fable 5 leads on agentic coding (SWE-bench Pro 80.3%) and backs it with safety classifiers, retention guarantees and Anthropic's enterprise support. This is the model's clearest win.

  • Retention and export-control matter to you

    Fable 5's mandatory 30-day retention, no-ZDR stance and availability outside export-controlled Chinese model channels make it the safer compliance pick for regulated enterprises. See the export-control section.

These are decision rules, not rankings — the right model depends on your workload, your data and your compliance posture.

Export control & compliance

A differentiator worth owning: US export-control rules and data-retention policy shape real deployment choices.

US export-control restrictions affect which advanced Chinese models are available and on what terms. As of August 2026, access and licensing for the cluster shift with the regulatory timeline — treat availability as a moving target and verify before you commit.

Fable 5's mandatory 30-day data retention and no-zero-data-retention (no-ZDR) stance are deliberate trade-offs for safety. Models that promise zero retention may not carry the same guarantees — a real difference if your data is sensitive.

Mainland-China access is a separate consideration: Chinese vendors often serve mainland users directly, while Anthropic's API is not available the same way in mainland China. Choose based on where your users and your data actually live.

Estimate the real cost

Headline token prices hide the workload cost — tokenizer inflation, caching, effort and output volume all move the number. Run the cost calculator for your actual usage before you pick.

Cost calculator

Which model matches your workload?

Compare across the full leaderboard, or price out your specific workload before you decide.

Frequently asked questions