Updated analysis
Fable 5 was suspended June 12–July 1, 2026 under US export controls and restored July 1; since July 24 Opus 5 is the Anthropic default. DeepSeek's open weights cannot be revoked by export control. Figures dated as of 2026-08-05.
See the alternatives guideFable 5 vs DeepSeek V4 Pro
Claude Fable 5 vs DeepSeek V4 Pro
The cheap open-weights flagship against Anthropic's closed flagship, as of August 2026 — with sourced benchmarks, retention and export-control reality, and a verdict by task.
TL;DR
DeepSeek V4 Pro is the cheap open-weights flagship: $0.435/$0.87 per million tokens (~23x cheaper than Fable 5 on input), 1M context, MIT weights you can self-host. Fable 5 wins decisively on agentic coding (SWE-bench Pro 80.3% vs self-reported 55.4%) and carries Anthropic's safety and mandatory-retention guarantees. Pick DeepSeek for cost, open weights and self-hosting; pick Fable 5 when coding quality and Anthropic guarantees dominate. As of 2026-08-05.
Specs at a glance
Fable 5 prices from our model pricing table (verified 2026-08-04); DeepSeek V4 Pro figures are 'as of' values from DeepSeek's official API docs.
| Spec | Claude Fable 5 | DeepSeek V4 Pro |
|---|---|---|
| Input price (per M tokens) | $10.00 | $0.435 |
| Output price (per M tokens) | $50.00 | $0.87 |
| Context window | 1M | 1M |
| Max output | 128K | 384K |
| Open source | Closed, API-only | MIT open weights (self-hostable) |
| Data retention | Mandatory 30 days (ZDR exempted) | Your policy (API) / yours entirely (self-host) |
Prices verified as of 2026-08-04: Fable 5 from the model pricing page, DeepSeek V4 Pro from DeepSeek's official API docs. DeepSeek has announced a future 2x peak-hour surcharge (Beijing peak hours) with no effective date yet — flat pricing applies as of 2026-08-05.
Benchmarks, with the honest source note
The two tables are not peers: Fable 5's numbers are Anthropic's official launch data; DeepSeek's are self-reported and not independently verified as of 2026-08-05.
| Spec | Claude Fable 5 | DeepSeek V4 Pro | Source |
|---|---|---|---|
| SWE-bench Pro (agentic coding) | 80.3% | 55.4% | Fable 5: Anthropic launch table (official). V4 Pro: DeepSeek technical report (self-reported) |
| FrontierCode (Diamond) | 29.3% | Not published | Anthropic launch table (official); DeepSeek not published |
| GDPval (knowledge work) | 1932 | Not published | Anthropic launch table (official); DeepSeek not published |
| OSWorld (computer use) | 85.0% | Not published | Anthropic launch table (official); DeepSeek not published |
Sources: Fable 5 figures from Anthropic's June 9, 2026 launch table (official). DeepSeek V4 Pro figures from DeepSeek's own technical report — self-reported, not independently verified as of 2026-08-05. Treat every DeepSeek row as directional.
Which one should you pick?
A verdict by task, not by model.
You want open weights or self-hosting - DeepSeek V4 Pro
MIT weights, 1M context, no per-token fee on your own hardware, and a model that cannot be switched off by export control. That permanence is not hypothetical for Fable 5 - it was suspended for 19 days.
You are cost-bound at scale - DeepSeek V4 Pro
At list prices V4 Pro is roughly 23x cheaper on input and 57x on output than Fable 5's $10/$50. For high-volume production traffic the multiple is the whole story.
Frontier agentic coding - Fable 5
80.3% on SWE-bench Pro is the highest published score for a public model, official and unrefuted. DeepSeek's 55.4% on the same test name is self-reported and unverified.
Compliance - Fable 5
Fable 5's mandatory 30-day retention rules it out for zero-retention or data-residency work - but open weights have no Anthropic safeguard gate either. If you need auditable, gated, retention-defined behavior, Fable 5 is the defensible choice.
Long-horizon autonomy - Fable 5
Fable 5 was built for multi-day agent runs at 1M context. DeepSeek's endurance at that scale is exactly what no one has independently verified.
There is one shared 'SWE-bench Pro' name (80.3% vs 55.4%), but the DeepSeek figure is its own harness run with zero third-party replication — same test name, different trust levels. Do not compare DeepSeek's SWE-bench Verified (80.6%) directly to Fable 5's SWE-bench Pro.
Cost: a worked example
A representative 10,000-session coding-agent workload shows how the list-price gap compounds in practice. Figures are indicative, as of 2026-08-05.
The same workload, roughly two orders of magnitude cheaper
On a typical coding-agent month (400M cached-heavy input tokens, 20M output, 80% prompt-cache hit), Fable 5 lands around $2,120 while DeepSeek V4 Pro comes to roughly $53 on its API - about 40x less. With no caching the gap narrows to ~26x, but it never flips.
Self-hosting removes the token fee entirely
With MIT weights you pay only for your own GPUs - no per-token fee, and your data never leaves your infrastructure. No Anthropic model offers this lane.
Run your exact token mix through the cost calculator - Fable 5's Batch API (50% off, not real-time) drops its side further, and DeepSeek publishes no batch tier.
Price your own workload
The benchmark gap and the price gap point different ways. Run your exact token mix through the cost calculator before you route, and keep a portable prompt across both models.