DeepSeek V4 Pro Closes on Claude Fable, Testing Anthropic Premium

A quiet upgrade with a loud comparison

DeepSeek did not announce its most pointed challenge to Anthropic. In mid-August, the API pricing page entry for "deepseek-v4-pro" changed to DeepSeek-V4-Pro-0813. That was the general-availability build of a model that had run as a preview since April, and it shipped without a blog post.41 DeepSeek's own comparison table did the talking. Across nine agent benchmarks where both models have scores, it places Anthropic's Claude Fable 5 ahead by an average of 5.3%, and DeepSeek wins outright on two of them.41 At launch, Fable 5 cost $10 per million input tokens and $50 per million output tokens, while V4 Pro cost $0.435 and $0.87.41

That combination produced the "5% better at 4,500% the price" framing, and the headline spread quickly. Two months later, the more useful questions are what the claim actually measured, what has changed since, and what it means for Anthropic's most expensive model tier. My reading is that the slogan overstated how close the models are but got the economics right. Anthropic's real problem is less a single Chinese model than the growing evidence that its top-tier price is hard to justify for routine work.

What the 5% figure does and does not say

The benchmark gap is narrower than it first appears, and also less reliable. Decrypt found that one row skews the average: Humanity's Last Exam without tools, where DeepSeek scores 42.7 to Fable's 53.3. Removing it brings Fable's average lead to about 2.8%.41 Every number in that table came from DeepSeek, though. When the 0813 build shipped, no outside lab had benchmarked it yet.41

Kingy AI's fact-check was more skeptical. It called the "Fable-level" claim partly supported but too broad. DeepSeek did report near-ties on Terminal Bench 2.1 and CyberGym, but the gaps were larger on HLE, DeepSWE, Toolathlon and DSBench-FullStack.42 The review also noted that DeepSeek's Fable column was labeled as a fallback configuration, and that the two models trade wins depending on which rows were chosen.42 Coverage that focused on the near-ties reported DeepSeek's CyberGym score of 83.3 against Fable 5's 83.1.43 Users on Anthropic's own subreddit highlighted a separate result where V4 Pro scored 87.9 to Fable's 88.0.44

Independent tests run since then put the gap somewhere between DeepSeek's chart and the skeptics' view. On Together AI's DeepSWE run, Fable 5 at maximum effort solved 69.7% of tasks and V4 Pro 0813 solved 62.8%. Fable won six of eight domains and four of five programming languages, including a 20-point lead in Rust. Artificial Analysis's composite index placed V4 Pro 0813 at 36 against Claude Opus 5 at roughly 51 in early October.17 On factual calibration, the gap is large: AA-Omniscience records 49.1% accuracy and a 94.8% hallucination rate for V4 Pro 0813, compared with 67.2% and 72.6% for Claude Fable 5.1.22

In short, "5% behind" describes a narrow set of agentic benchmarks chosen by the vendor. It does not hold across the board.

The price gap is real, but it's shrinking

On cost, the coverage mostly agrees, though the size of the multiple depends on how it is calculated. Kingy calculated that the widely quoted 57x figure holds only for output tokens. Base input was closer to 23x, and a simple equal mix of input and output came to about 46x.42 That last figure is roughly where the "4,500%" headline and Crypto Briefing's "45 times" framing came from.46

Per task, the gap is larger. Artificial Analysis measured Fable 5 at $3.15 per benchmark task, because Fable reasons longer and writes more.41 Together AI's DeepSWE runs cost $21.63 per rollout on Fable versus $0.24 on V4 Pro, about 90 times cheaper. That works out to 260 solved tasks per $100 for V4 Pro and three for Fable.

The launch prices did not last. Kingy pointed out at the time that DeepSeek had warned of significant price increases.42 Peak and off-peak billing started on August 16. V4 Pro now costs $0.66 input and $1.98 output per million tokens off-peak, and twice that at peak.14 At peak, Fable's $50 output rate is about 12.6 times V4 Pro's $3.96, not 57 times.314 One guide claims V4 Pro traffic has been billed at Flash rates since September 14.40 Other trackers report that DeepSeek dropped that rerouting plan and kept V4 Pro at its own prices.3134 The later reporting looks more reliable to me.

Even at 12x, the gap is large. It is no longer an order-of-magnitude story, though, and the narrowing undercuts the idea that DeepSeek can keep pricing near-frontier models almost at cost indefinitely.

The Anthropic angle: the challenge comes from inside the lineup too

The most interesting point in the original Decrypt piece was about Anthropic's own models: Claude Opus 5 outscored Fable 5 on most benchmarks at half the price.41 That has become more pronounced since. Anthropic's documentation for Fable 5.1, released September 1, tells developers to start most work on Claude Opus 5.5, which costs $4 input and $20 output. Fable is reserved for long-horizon agentic work, or for cases where Opus at higher effort still falls short.3 Toloka reports that Anthropic says Opus 5.5 matches Fable 5.1 on most tasks, and on Toloka Arena's multi-turn tool-use tasks Opus 5.5 scores 77.3 against Fable 5.1's 74.0.9

In other words, Anthropic itself does not present Fable as the default for everyday work. It presents Fable as a premium option for long, autonomous jobs. Its marketing makes the same case: the company says Fable's lead over competitors grows with task length and complexity.7 Jane Street told Anthropic that Fable 5.1 stays readable over long, multi-step tasks, which earlier models struggled with.1 A head-to-head benchmark comparison with DeepSeek mostly misses that argument, because short, vendor-chosen agent benchmarks are where cheap models look strongest.

Anthropic has also responded on price, though without cutting its list rates. Fable 5.1 kept Fable 5's $10/$50 pricing but cut cache-read costs to a quarter.3 Cognition, which makes the Devin coding agent, said that change made a Fable-class model affordable for work it had kept on Opus.1 For long agent loops dominated by cached context, that is the price that matters most, and it is exactly where DeepSeek is cheapest: $0.022 per million cached tokens off-peak.15

Trust and access are part of the price

The comparison is not only about benchmarks and tokens. In June, Fable 5 had a rough launch: the U.S. government barred non-U.S. nationals from accessing it and Mythos 5, Anthropic pulled access entirely, and service resumed on July 1 after Commerce lifted the restrictions.6 Fable's safeguards also affect behavior. When its classifiers flag requests about cybersecurity, biology and chemistry, or model distillation, the request is handed to a less capable Opus model.6 One developer analysis criticized this as silent degradation for anyone suspected of building a competing AI.7 Another alternatives guide argued that Fable's 30-day data retention is a dealbreaker for some teams, and that self-hosting open-weight DeepSeek models removes the problem.21

DeepSeek does not offer the same guarantees, but it offers control. V4 Pro is a 1.6-trillion-parameter mixture-of-experts model with 49 billion active parameters per token, and its weights are released under the MIT license.26 Kingy noted at launch that DeepSeek had not confirmed whether the downloadable weights matched the 0813 API build.42

What buyers are actually doing

The most practical evidence points to routing, not replacing one model with the other. Together AI found that sending DeepSWE tasks to V4 Pro first, and passing only failures to Fable, solved 82.7% of tasks at $8.28 each. Fable alone solved 69.7% at $21.63. The two models fail on different tasks, so pairing them beats either one alone, and Fable is paid only for the hardest problems.

For Anthropic, that is a manageable outcome. Its premium tier stays in use for the hardest work, but the volume beneath it goes to cheaper models, whether DeepSeek's or Anthropic's own Opus. DeepSeek's position is also changing. Its V4.1 Flash, released September 10, beats V4 Pro on every agentic coding, cyber and tool-use row DeepSeek reports.34 As of early October, no V4.1 Pro had shipped.31

The August headline was right that Chinese open-weight models are getting within a few points of the U.S. frontier on selected benchmarks for a fraction of the price.41 It overstated how close they are on reliability, calibration and very long tasks. For Anthropic, the takeaway is that Fable's $50 output price is now justified by those long, difficult jobs, and the company has already started steering everything else to cheaper models.