Verdict: DeepSeek V4 Pro 0813 is the current API version, and DeepSeek’s launch table puts it close to Claude Fable 5 on a few agent benchmarks. But “Fable-level at 1/57 the price” is a launch-era slogan, not the current price. DeepSeek’s comparison chart is vendor-reported. The August 13 Artificial Analysis snapshot supports a higher index score than April’s V4 Pro baseline, but it does not establish broad Fable 5 equivalence. At the rates checked October 7, DeepSeek output is about 1/13 of Fable 5 at peak pricing or 1/25 off peak. An October 8 capture of DeepSeek’s official 0813 model card identifies the weights as the official 0813 release.
This is a breaking-news source audit, not a definitive hands-on review. Kingy did not run controlled prompts, measure latency, inspect API outputs, or reproduce the benchmark harnesses. The August 13 Artificial Analysis snapshot below supports a higher composite score than its April V4 Pro baseline, but does not establish broad Fable 5 equivalence or reproduce DeepSeek’s selected agent chart. The official 0813 model card captured October 8 now establishes that DeepSeek published 0813 weights; the article does not test local deployment.
September 5 check (historical): DeepSeek’s official changelog documented the August 13 general-availability rollout across its app, web and API, and said the revised price schedule took effect August 16. The pricing page identified 0813 as the current API version. At that check, the general V4 Pro repository had no 0813 tag or revision, so the relationship between those files and the API checkpoint remained unconfirmed. An October 8 official 0813 model card now resolves the weights question; neither source validates every row in DeepSeek’s launch chart.
| Claim | Fact-check | Why |
|---|---|---|
| “0813 is live” | Supported for API access | DeepSeek’s August 13 changelog documents general availability, and its pricing page names DeepSeek-V4-Pro-0813 as the current version. |
| “Fable-level” | Partly supported, too broad as a blanket claim | DeepSeek reports near-ties on Terminal Bench 2.1 and CyberGym, but larger gaps on HLE, DeepSWE, Toolathlon and DSBench-FullStack. |
| “1/57 the price” | Launch-era claim; no longer current | At the October 7 rate check, output was $3.96 per million at peak or $1.98 off peak, versus Fable 5 at $50: about 1/13 or 1/25. |
| “0813 weights are available” | Confirmed in a later source | The September 5 check had not identified 0813 weights. DeepSeek’s official 0813 model card, captured October 8, calls itself the official release and lists MIT-licensed model weights. |
Commercial relationship for this article: None. This is based on Curtis Pyke’s direct confirmation on October 7, 2026.
What Is Actually Live?
DeepSeek’s official Models & Pricing page now identifies the current deepseek-v4-pro version as DeepSeek-V4-Pro-0813. The API alias stays the same. The same page lists a one-million-token context window, a maximum output of 384,000 tokens, thinking and non-thinking modes, tool calls, JSON output, the Responses API and Anthropic-compatible access. Artificial Analysis and OpenRouter both list the route as text-to-text; neither record advertises image input.
DeepSeek’s August 13 changelog entry confirms the general-availability rollout across the app, web and API. It also says the API alias remains unchanged, so callers use deepseek-v4-pro while the pricing page identifies the current underlying version as 0813.
The September 5 check established API availability; the later-dated public-weight evidence is assessed below and does not change what was known then.
What the Fuller Official Specs Add
DeepSeek’s official repository describes V4 Pro as a mixture-of-experts model with 1.6 trillion total parameters and 49 billion activated per token, trained as part of the V4 family on more than 32 trillion tokens. At a one-million-token context, DeepSeek says the architecture needs 27% of V3.2’s single-token inference FLOPs and 10% of its KV-cache footprint. These are DeepSeek-reported architecture and training statistics for the V4 release family; the repository still does not identify its files or model-card evaluations as the 0813 API revision.
DeepSeek’s August 13 update also makes benchmark configuration more consequential than a single launch score suggests: 0813 supports low, high and max thinking effort. The older repository results below use non-think, high and max for the earlier V4 Pro family, not a documented 0813 evaluation setup. Evaluation claims therefore need the model revision, effort setting, harness, benchmark version and token budget.
OpenRouter adds one useful live-serving detail: its model record says 0813 is currently hosted by one provider and requests are forwarded directly to DeepSeek. Its public endpoint record currently leaves latency and throughput fields empty. That absence is not evidence of slow service; it means OpenRouter does not yet provide a durable serving measurement worth citing.
Reasoning Effort Can Move the Reported Scores Sharply
The official V4 model card publishes mode-by-mode results for the earlier V4 Pro release family. They are material because they show how strongly scores can depend on reasoning budget. They are not 0813 evaluations, and they use different benchmark versions and harnesses from the new agent chart, so the two tables should not be merged.
| DeepSeek-reported repository eval | V4 Pro non-think | V4 Pro high | V4 Pro max | Non-think to max |
|---|---|---|---|---|
| LiveCodeBench (Pass@1) | 56.8 | 89.8 | 93.5 | +36.7 points |
| HLE (Pass@1) | 7.7 | 34.5 | 37.7 | +30.0 points |
| MRCR 1M (MMR) | 44.7 | 83.3 | 83.5 | +38.8 points |
| SWE-bench Verified (resolved) | 73.6 | 79.4 | 80.6 | +7.0 points |
The practical conclusion is narrower than “max is always better.” In DeepSeek’s own table, moving from high to max adds 3.7 points on LiveCodeBench and 3.2 on HLE, but only 0.2 on MRCR 1M and 1.2 on SWE-bench Verified. Evaluation claims therefore need the model revision, effort setting, harness, benchmark version and token budget. DeepSeek’s official 0813 model card, captured October 8, adds that code-agent tasks among its public benchmarks used the minimal mode of DeepSeek Harness, max reasoning effort, temperature 1.0 and top_p 0.95. Those vendor-reported settings improve the description of its method; Kingy has not independently reproduced the chart, and the card does not give enough detail to reproduce every benchmark result.
The First Independent Result Has Arrived
Artificial Analysis has a dedicated V4 Pro 0813 page and describes its measurement as an independent run on dedicated hardware. The values below are its August 13, 2026 snapshot, not a current-score claim. At max reasoning effort, 0813 scores 53 on Artificial Analysis Intelligence Index v4.1.1, versus 45 for the April V4 Pro baseline. The v4.1.1 index at that time combined nine evaluations, including Terminal-Bench 2.1, Humanity’s Last Exam, GPQA Diamond, SciCode, GDPval-AA v2 and long-context and agentic components.
| Independent Artificial Analysis measure | V4 Pro April | V4 Pro 0813 | Change |
|---|---|---|---|
| Intelligence Index | 45 | 53 | +8 points |
| Output speed | 70.6 tokens/s | 83.2 tokens/s | About 18% faster |
| Time to first token | 1.68s | 1.63s | 0.05s faster |
| Output tokens used by the index | 180M | 130M | About 28% fewer |
| Total index evaluation cost | $176.25 | $135.03 | About 23% lower |
Later snapshot: In the October 8 capture, this Artificial Analysis page displayed score 36 on Intelligence Index v4.3.2, whose page lists 10 evaluations. That is a different index version and evaluation set from the August 13 v4.1.1 snapshot above (score 53, nine evaluations). Keep the dated results separate; they do not show a comparable score change.
This August 13 snapshot independently supports a higher index score and lower index token use than the April route on that version of Artificial Analysis’s evaluation. It still does not reproduce DeepSeek’s Fable comparison: Artificial Analysis uses its own composite, settings and graders, and the public model page does not turn a score of 53 into a blanket Fable 5 equivalence claim.
The Benchmark Chart Is DeepSeek-Reported
The supplied benchmark image carries DeepSeek branding and compares V4 Pro 0813 with V4 Flash 0731, the two V4 previews, GLM-5.2, Kimi-K3, Opus 4.8 and “Fable 5 (w/ fallback).” Those numbers should be read as DeepSeek-reported launch results, not Kingy measurements and not an independent leaderboard.
| Benchmark | DeepSeek V4 Pro 0813 | Fable 5 (with fallback) | Difference |
|---|---|---|---|
| HLE without / with tools | 42.7 / 60.0 | 53.3 / 63.0 | DeepSeek trails by 10.6 / 3.0 points |
| Terminal Bench 2.1 | 87.9 | 88.0 | DeepSeek trails by 0.1 point |
| CyberGym | 83.3 | 83.1 | DeepSeek leads by 0.2 point |
| DeepSWE | 62.7 | 70.0 | DeepSeek trails by 7.3 points |
| Toolathlon-Verified | 74.1 | 77.9 | DeepSeek trails by 3.8 points |
| AutomationBench (Public) | 31.8 | 29.1 | DeepSeek leads by 2.7 points |
| DSBench-FullStack | 71.1 | 77.2 | DeepSeek trails by 6.1 points |
| DSBench-Hard | 67.2 | 68.3 | DeepSeek trails by 1.1 points |
The honest reading is Fable-adjacent on parts of this selected agent suite. DeepSeek is effectively tied on Terminal Bench 2.1 and CyberGym, leads on AutomationBench, and comes within 1.1 points on DSBench-Hard. It is not tied on every row: Fable leads by 7.3 points on DeepSWE, 6.1 on DSBench-FullStack and 10.6 on HLE without tools.
The fallback label also matters. Anthropic’s official Fable 5 documentation says queries flagged by its cybersecurity and biology safeguards may be routed to less capable models. DeepSeek’s comparison therefore describes Anthropic’s deployed Fable-with-fallback product behavior on this chart; it does not establish equivalence to every possible Fable configuration or workload.
The Percentage-Point Correction
Fresh X posts repeated the preview-to-0813 jumps, sometimes presenting the scores as percentages. The subtraction is measured in percentage points when both values are rates on a 0–100 scale. Relative percentage growth is a different calculation.
| DeepSeek-reported benchmark | V4 Pro Preview | V4 Pro 0813 | Correct point gain | Relative increase |
|---|---|---|---|---|
| Terminal Bench 2.1 | 72.1 | 87.9 | +15.8 points | +21.9% |
| CyberGym | 52.7 | 83.3 | +30.6 points | +58.1% |
| DeepSWE | 12.8 | 62.7 | +49.9 points | +389.8% |
| AutomationBench (Public) | 12.8 | 31.8 | +19.0 points | +148.4% |
| DSBench-Hard | 31.1 | 67.2 | +36.1 points | +116.1% |
Calling DeepSWE a “49.9% improvement” would understate the relative change and misuse the unit. The score rose 49.9 percentage points; relative to the 12.8 starting value, the increase is about 390%. Neither number proves real-world task completion without the harness, settings, sample sizes and reproducibility data.
Is the Launch-Era 1/57 Price Claim Still True?
No. DeepSeek’s pricing page, checked October 7, lists $1.32 per million cache-miss input tokens and $3.96 per million output tokens at peak, or $0.66 and $1.98 off peak. Anthropic lists Fable 5 at $10 per million base input tokens and $50 per million output tokens. The old 1/57 output comparison described the launch rate, not the current schedule.
| Billing component | DeepSeek V4 Pro 0813 | Claude Fable 5 | DeepSeek share of Fable price |
|---|---|---|---|
| Base / cache-miss input | $1.32 peak / $0.66 off peak per 1M | $10 / 1M | About 1/8 peak / 1/15 off peak |
| Cache read / hit input | $0.044 peak / $0.022 off peak per 1M | $1 / 1M | About 1/23 peak / 1/45 off peak |
| Output | $3.96 peak / $1.98 off peak per 1M | $50 / 1M | About 1/13 peak / 1/25 off peak |
| Simple 1M input + 1M output example | $5.28 peak / $2.64 off peak | $60 | About 1/11 peak / 1/23 off peak |
Real agent bills depend on the input/output mix, reasoning-token treatment, cache hits, retries and tool-loop length. Fable’s prompt cache writes also have their own rates: Anthropic lists $12.50 per million tokens for a five-minute cache write and $20 for a one-hour write, then $1 for cache reads. DeepSeek defines weekday peak hours as 01:00–04:00 and 06:00–10:00 UTC and says other times are off peak; buyers should check the official rate card before estimating a workload.
KV Cache Is Not the Same as Prompt Cache Pricing
The official DeepSeek V4 repository says V4 Pro requires 10% of DeepSeek V3.2’s KV-cache footprint at a one-million-token context. That is an inference-memory efficiency claim: the key-value cache stores attention state while a model processes and generates from a long sequence.
DeepSeek’s API context-caching feature is a separate customer-facing mechanism. It stores reusable prompt prefixes on disk, is enabled by default, and bills matching input tokens at the cache-hit rate. DeepSeek calls it best-effort and does not guarantee every repeated prefix will hit.
Anthropic’s prompt caching likewise reuses prompt prefixes, but exposes explicit or automatic cache controls, five-minute or one-hour lifetimes, and separate write/read prices. These concepts are related inside inference systems, but the numbers are not interchangeable. “90% less KV cache” does not mean “90% off the API bill,” and a DeepSeek cache-hit price should not be compared with a Fable base-input price as if every request automatically qualifies on both sides.
Are the 0813 Open Weights Available?
In the September 5 check below, the general V4 Pro repository had no August 0813 tag or revision, so the link between its files and the API checkpoint was then unconfirmed. An October 8 capture of DeepSeek’s separate official V4 Pro 0813 model card now calls 0813 the official release, marks the repository and model weights MIT-licensed, and gives vLLM and SGLang instructions. That resolves whether DeepSeek has published official 0813 weights. It does not establish that a particular machine can serve this 1.7-trillion-parameter model, or that its local behavior matches the hosted API.
What the Audited X Posts Got Right—and What They Overstated
- ChrisGPT correctly attributed the benchmark jumps to DeepSeek and quoted the launch-era price and context window. The post also noted the problem of companies reporting different benchmark sets.
- DanDr1s accurately reproduced five rows from the supplied DeepSeek chart and the launch-era $0.435/$0.87 rates. Those are not the current schedule. The gains should be described as points, not percentages.
- Junyuan Shang called the result “nearly Fable 5-level.” That is defensible for a subset of the chart, but too broad without the rows where Fable leads and without independent replication.
- Chizzydigital paired the Fable-level framing with “access to the weights.” DeepSeek’s official 0813 model card captured October 8 now supports the weights claim; it does not establish what was available when the post appeared.
Kingy’s Breaking-News Read
DeepSeek V4 Pro 0813 is one of the most aggressive price-performance launches of 2026. Artificial Analysis’s August 13 score of 53—up from 45 for April V4 Pro on index v4.1.1—was a higher result on that dated composite, while its lower token use and evaluation cost strengthened that snapshot’s price-performance case. Its October 8 page uses a different index version and evaluation set, so the dated scores do not establish a trend. DeepSeek’s near-parity claim with Fable 5 on Terminal Bench 2.1 and CyberGym remains vendor-reported.
But the evidence does not justify a universal “Fable for 1/57 the cost” verdict. DeepSeek’s chart is vendor-reported, its Fable column is explicitly a fallback configuration, the models trade wins across the selected rows, and the launch-era 57-times figure is not the current rate. The evidence-bounded conclusion is: 0813 is a live, comparatively inexpensive text-only API model; its August 13 Artificial Analysis result was higher than the April baseline on that index version, and DeepSeek’s October 8 official model card identifies published 0813 weights. DeepSeek’s selected Fable-adjacent chart remains vendor-reported and unreproduced here.
FAQ
Is DeepSeek V4 Pro 0813 live?
Yes. DeepSeek’s official August 13 changelog documents general availability across its app, web and API, and its pricing page identifies the current deepseek-v4-pro version as DeepSeek-V4-Pro-0813.
Does DeepSeek V4 Pro 0813 match Claude Fable 5?
DeepSeek reports near-ties on some agent benchmarks, including Terminal Bench 2.1 and CyberGym. Fable leads by wider margins on other rows such as DeepSWE, HLE without tools and DSBench-FullStack. No independent Kingy test has established broad equivalence.
Is DeepSeek really 57 times cheaper than Fable 5?
No, not at current rates. Output is $3.96 per million at peak or $1.98 off peak versus Fable 5 at $50, or about 1/13 and 1/25 respectively. The total ratio depends on peak timing, token mix and cache behavior.
Are the DeepSeek V4 Pro 0813 weights open?
Yes. In an October 8 capture, DeepSeek’s official V4 Pro 0813 model card calls 0813 its official release and lists the model weights under MIT. That later evidence supersedes the September 5 repository check; it does not establish that a particular machine can run the 1.7-trillion-parameter model.
Does DeepSeek V4 Pro 0813 support images?
No image input is listed in the current Artificial Analysis or OpenRouter records; both describe 0813 as text-to-text. Treat the API model as text-only unless DeepSeek documents a vision capability.
What is the difference between KV cache and prompt caching?
KV cache is the model’s attention-state memory during inference. Prompt caching is an API feature that reuses matching input prefixes across requests and can change billed input cost. A reduction in KV-cache memory does not directly equal the prompt-cache discount.
Sources
- DeepSeek API: Models & Pricing
- DeepSeek API: Thinking Mode and effort mapping
- DeepSeek API: Rate Limit & Isolation
- DeepSeek API changelog
- DeepSeek API: Context Caching
- Official DeepSeek V4 Pro repository and model card
- Official DeepSeek V4 Pro 0813 model card and weights (captured October 8)
- Official DeepSeek V4 Pro commit history
- OpenRouter model record: DeepSeek V4 Pro 0813
- OpenRouter launch post on X
- Anthropic: Claude Fable 5
- Anthropic: Prompt caching and Fable 5 cache pricing
- Artificial Analysis: DeepSeek V4 Pro 0813 independent evaluation
- Artificial Analysis: April V4 Pro baseline
- Artificial Analysis Intelligence Index methodology
- Audited X post: ChrisGPT
- Audited X post: DanDr1s
- Audited X post: Junyuan Shang
- Audited X post: Chizzydigital