Comparison · same table

Pro 0813 vs Fable 5
Tied on TB2.1, Fable still ahead on DeepSWE

On DeepSeek's own 0813 table, Terminal-Bench 2.1 reads 87.9 vs 88.0 — essentially a tie, with the Fable 5 column labelled w/ fallback. Price is the real divide: $0.435/$0.87 against $10/$50.

#TB2.1 87.9 vs 88.0#DeepSWE 62.7 vs 70#HLE 60 vs 63#Open vs closed

Highlights

87.9 vs 88.0

TB 2.1

Essentially tied. These are Terminal-Bench 2.1 numbers; Grok's launch table uses TB v3.0, a different scale that is not directly comparable.

62.7 vs 70.0

DeepSWE

Fable 5 (w/ fallback) still leads the hard coding-agent row.

60.0 vs 63.0

HLE w/ tools

Fable 53.3/63.0 vs Pro 42.7/60.0 (wo/w).

$0.87 vs site

Output price

V4 Pro output is $0.87/M today; Claude Fable 5 lists officially at $10/M input and $50/M output.

The matchup

This is the search that asks if 0813 is a Fable substitute. On DeepSeek’s own grid it is a peer on terminal-bench and still behind on DeepSWE, Toolathlon and internal fullstack.

Why it is trending

The community used April's V4 Pro Preview as a punchline on Baba-is-bench. Those 29% figures belong to the Preview build; the 0813 release has no published score on that benchmark.

Timeline

2026-08-13

Pro 0813 GA + HF comparison table including Fable-5 w/ fallback.

Fable 5 remains Anthropic’s premium story; price lives on the existing QCode Fable pages.

2026-08-16

DeepSeek peak/off-peak. Still far under Fable-class API bills.

How to quote the HF grid

On the table

The table below uses only DeepSeek's same-source 0813 grid published on Hugging Face, where the Fable 5 column is labelled w/ fallback. Cybergym 83.3 vs 83.1 is a fair tie.

Off the table

Three common misreadings: treating blog rumours as Fable 5's official price, pinning the Preview build's low scores on 0813, and putting Terminal-Bench 2.1 and v3.0 numbers in the same table.

Who should pick what

Need Fable judgment / brand / Claude stack

Stay on Fable 5 or Opus 5. The HF grid still gives Fable the harder agent rows.

Need open 1.7T and a 30–50× cheaper token bill

Pro 0813. Self-host or official API. Pair with Flash 0731 for volume.

How to use it

Same 1M context. Different products: one MIT checkpoint, one closed premium model.

On QCode

QCode currently lists both deepseek-v4-pro and claude-fable-5 on /models. QCode is a multi-model access layer and is not affiliated with either vendor.

FAQ

Is 0813 equal to Fable 5?

No. Tied on TB2.1; behind on DeepSWE and several agent rows.

Why w/ fallback on Fable?

That is how DeepSeek labeled the column. Omitting it overstates Fable or understates Pro.

Can I self-host the Fable side?

No. Only Pro 0813 is MIT.

Flash 0731 vs this matchup?

Flash is cheaper and close on some agent rows; 0813 is the Pro-side of the Fable question.

Old 4.8 Opus in the HF table?

DeepSeek included Opus-4.8 as a published reference, not as this page’s hero opponent.

Peak price kill the gap?

No. Even peak Pro output $3.96 is not Fable-class billing.

Sources

HF DeepSeek-V4-Pro-0813 table; DeepSeek changelog 2026-08-13; site Fable pricing pages.

Peer on terminal, not a Fable clone

Keep Fable for the last 5% of hard agent work; use 0813 when the bill is the product.