---
type: analysis
slug: research-deepseek-v4-pro-0813-is-real-cheaper-than-fable-5-and-not-a-silent-57x-clone
title: "DeepSeek V4-Pro-0813 is real, cheaper than Fable 5, and not a silent 57x clone"
description: "DeepSeek published V4-Pro-0813 as the official Pro GA on 13 August 2026 with the 1.6T, 49B-active, 1M-context spec Cline advertised. Vendor Terminal Bench gains and a roughly 57 times cheaper Fable slogan do not survive current list prices or independent evals."
published_at: 2026-10-26T12:00:00.000Z
topics: ["evals","research"]
source_urls: ["https://api-docs.deepseek.com/updates/","https://api-docs.deepseek.com/news/news260813","https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813","https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813/commits/main","https://www.deepseek.com/en/news/v4-preview/","https://fe-static.deepseek.com/chat/transparency/deepseek-V4-model-card-EN.pdf","https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro","https://catalog.ngc.nvidia.com/orgs/nim/deepseek-ai/models/deepseek-v4-pro-0813/-","https://api-docs.deepseek.com/quick_start/pricing/","https://artificialanalysis.ai/models/deepseek-v4-pro","https://www.tbench.ai/leaderboard/terminal-bench/2.1?agents=Terminus+2","https://platform.claude.com/docs/en/about-claude/pricing","https://platform.claude.com/docs/en/models/fable-5/overview","https://the-decoder.com/deepseek-launches-an-improved-v4-pro-model-raises-api-prices-and-makes-its-agent-software-open-source/","https://artificialanalysis.ai/models/releases/comparisons/gpt-6-astra-vs-deepseek-v4-pro","https://artificialanalysis.ai/evaluations/terminalbench-2-1?models=deepseek-v4-pro","https://docs.cline.bot/getting-started/clinepass","https://cline.bot/models/deepseek-v4-pro","https://api-docs.deepseek.com/news/news260910"]
author: "Andrey Reshetnikov"
author_profile: https://newruntime.com/owner-profile.md
---
Cline promoted DeepSeek-V4-Pro-0813 around 12–13 August 2026 as a silent drop: 15.8 percent better on Terminal Bench than the April preview, Fable 5-level agent performance at about 57 times lower cost, 1.6 trillion total parameters with 49 billion active and a 1 million-token context, plus availability in ClinePass. That pitch is an integrator post, not a model card, which is why a news item alone could not settle it.

The checkpoint still sits behind DeepSeek’s deepseek-v4-pro API as of 23 October 2026. MIT-licensed weights are on Hugging Face, and NVIDIA ships a NIM artifact. Builders deciding whether this is a Fable-class agent at commodity prices need a split between what DeepSeek shipped and what Cline claimed. Primary docs confirm the GA, the size, and the context window. Current list prices, DeepSeek’s own comparison table, and Artificial Analysis scores cut against silent, Fable-equivalent, 57 times cheaper at today’s rates, and best on the market.

## What we checked

- **DeepSeek silently released V4-Pro 0813 as an unofficial checkpoint.** partly confirmed. Weights for DeepSeek-V4-Pro-0813 landed on Hugging Face on 13 August 2026, which can look like a quiet drop. The same calendar day DeepSeek posted an API changelog and a GA news item stating the Pro endpoint, app, and web Expert Mode now serve that build. [3][4][1][2]
- **V4-Pro-0813 is up 15.8% on Terminal Bench versus the April Preview.** partly confirmed. DeepSeek’s 0813 table reports Terminal Bench 2.1 at 87.9 versus 72.1 for V4-Pro Preview, a 15.8-point gap that matches Cline’s figure. That is a vendor score in DeepSeek Harness minimal mode at max effort, not a relative percent on the public Terminus leaderboard. [3][1][7][11]
- **0813 reaches Fable 5 performance at about 57 times lower cost.** partly confirmed. On DeepSeek’s own table, 0813 is near Fable-5 with fallback on Terminal Bench 2.1 and Cybergym and behind on HLE, DeepSWE, and Toolathlon. Current list prices are $1.32/$3.96 peak for Pro versus $10/$50 for Fable 5, about 13 times cheaper on peak output, not 57; the 57 times figure matches Fable’s $50 output divided by DeepSeek’s pre-16-August $0.87 output rate. [3][9][12][13][14]
- **The model is 1.6T parameters, 49B active, with 1M context.** confirmed. DeepSeek’s April preview blog, April English model card, preview Hugging Face card, NVIDIA NGC listing, and Artificial Analysis all state 1.6T total, 49B active, and 1M context. The live pricing page still names the Pro SKU DeepSeek-V4-Pro-0813 with 1M context and 384K max output. NVIDIA prints 1.65T total, a rounding difference. [5][6][7][8][9][10]
- **This is the best price-to-performance model on the market.** refuted. No primary source ranks 0813 as the market Pareto leader. Artificial Analysis currently scores it 36 on Intelligence Index v4.3.2 and 14.1% on Terminal-Bench 4.0, far below GPT-6 Astra (max) on those tests, while DeepSeek later said V4.1-Flash beat V4-Pro on performance, cost, speed, and runtime. [10][15][19]
- **V4-Pro-0813 is available in ClinePass.** confirmed. ClinePass documentation lists DeepSeek V4 Pro as cline-pass/deepseek-v4-pro and Cline’s model page marks it available in ClinePass. The marketing page was last reviewed 14 July 2026, before 0813, so the docs confirm Pro access, not a Cline-run 0813 eval. [17][18]

## What we found

- DeepSeek documented 0813 the same day the weights appeared. The 13 August 2026 changelog calls it the GA rollout on app, web, and API under the unchanged name deepseek-v4-pro, lists agent scores including Terminal Bench 2.1 at 87.9, and links a GA news post. Hugging Face commits for deepseek-ai/DeepSeek-V4-Pro-0813 are dated 13 August 2026. Cline’s silent framing fits a same-day upload, not an unofficial model. [1][2][3][4]
- The 1.6T / 49B / 1M spec is consistent across DeepSeek’s 24 April 2026 preview blog, the April English model card PDF, the preview Hugging Face card, NVIDIA NGC, and Artificial Analysis. The live Models and Pricing page still names the Pro version DeepSeek-V4-Pro-0813, with 1M context and 384K max output. NVIDIA lists 1.65T total parameters, a rounding difference from DeepSeek’s 1.6T. [5][6][7][8][9][10]
- Cline’s 15.8 Terminal Bench figure is the point gap on DeepSeek’s table (87.9 versus 72.1), not a public-harness percent. The 0813 card says code-agent public benches used DeepSeek Harness minimal mode at max reasoning effort, temperature 1.0, top_p 0.95. The April instruct table reported Terminal Bench 2.0 at 67.9 for Pro-Max, so the baseline version also moved. The official Terminal-Bench 2.1 leaderboard page fetched here did not display a DeepSeek 0813 row. [3][7][1][11]
- Fable 5 parity is a selected vendor table, not a product equivalence. DeepSeek is near Fable-5 (w/ fallback) on Terminal Bench 2.1 (87.9 vs 88.0) and Cybergym (83.3 vs 83.1) and behind on HLE (42.7/60.0 vs 53.3/63.0), DeepSWE (62.7 vs 70.0), and Toolathlon-Verified (74.1 vs 77.9). Anthropic lists Fable 5 at $10 / $50 per million tokens. DeepSeek’s live Pro rates are $1.32 / $3.96 peak and $0.66 / $1.98 off-peak. Peak output is about 13 times cheaper, not 57. The Decoder reported the pre-16-August Pro output list price as $0.87; $50 divided by $0.87 is about 57. [3][12][13][9][14]
- Independent evals show a cheaper, mid-pack agent, not a Fable-class win. Artificial Analysis currently scores DeepSeek V4 Pro 0813 (max) 36 on Intelligence Index v4.3.2, 14.1% on Terminal-Bench 4.0, and $0.67 per index task, versus GPT-6 Astra (max) at 53, 59.1%, and $3.26. The Decoder said the older AA index rose from 45 to 53 on launch day, still behind Kimi K3 and Qwen 3.8 Max. AA’s Terminal-Bench 2.1 methodology uses Terminus 2, but the fetched page did not expose a numeric DeepSeek 0813 score. [10][15][14][16]
- ClinePass still offers DeepSeek V4 Pro, while DeepSeek’s own stack has moved. ClinePass docs list cline-pass/deepseek-v4-pro at $9.99 per month with 2–5 times usage versus standard API rates. On 10 September 2026 DeepSeek said tests put V4.1-Flash ahead of V4-Pro and that Pro would be phased out; the 17 October changelog says Pro API service continued after 14 September, and pricing still maps deepseek-v4-pro to 0813. [17][18][19][1][9]

## What it means

Treat V4-Pro-0813 as a real, MIT-licensed Pro GA with 1M context and a live deepseek-v4-pro alias, not as a stealth Fable substitute. If you need that checkpoint, pin the Hugging Face revision or the named Pro SKU and budget at current peak and off-peak rates, including cache-hit costs that The Decoder noted rose sharply on 16 August 2026. Reproduce agent quality in your harness; DeepSeek’s 87.9 used DeepSeek Harness at max effort, and independent Terminal-Bench 4.0 is 14.1% for 0813 max.

Do not freeze an architecture on Cline’s August slogan. DeepSeek later said V4.1-Flash beat Pro on performance, cost, speed, and runtime, then kept Pro only after demand. A/B Flash versus Pro on your traces, and remember ClinePass is a quota product, not an independent leaderboard. Self-host only if you can run the hybrid-attention MoE plus the attached DSpark draft module that NVIDIA and the 0813 card describe.

## Still open

- What numeric Terminal-Bench 2.1 score does Artificial Analysis or the official Terminus 2 leaderboard assign to DeepSeek-V4-Pro-0813, as opposed to DeepSeek’s 87.9 in DeepSeek Harness?
- Does ClinePass deepseek-v4-pro route to the 0813 weights, a later Pro, or a gateway mix, given the model page was last reviewed 14 July 2026?
- How much of the vendor agent lift is the 0813 checkpoint versus DeepSeek Harness minimal mode at max reasoning effort?
- How were Fable-5 with fallback scores collected, and do they match Anthropic’s current Fable 5.1 default-fallback product?

## Sources

1. [Change Log | DeepSeek API Docs](https://api-docs.deepseek.com/updates/) (primary)
2. [DeepSeek-V4-Pro GA Release | DeepSeek API Docs](https://api-docs.deepseek.com/news/news260813) (primary)
3. [deepseek-ai/DeepSeek-V4-Pro-0813 · Hugging Face](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813) (primary)
4. [Commits · deepseek-ai/DeepSeek-V4-Pro-0813](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813/commits/main) (primary)
5. [DeepSeek-V4 Preview: Entering the Era of Affordable Million-Token Context](https://www.deepseek.com/en/news/v4-preview/) (primary)
6. [DeepSeek V4 Technical Documentation (English model card PDF)](https://fe-static.deepseek.com/chat/transparency/deepseek-V4-model-card-EN.pdf) (primary)
7. [deepseek-ai/DeepSeek-V4-Pro · Hugging Face](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro) (primary)
8. [DeepSeek-V4-Pro-0813 | NVIDIA NGC](https://catalog.ngc.nvidia.com/orgs/nim/deepseek-ai/models/deepseek-v4-pro-0813/-)
9. [Models & Pricing | DeepSeek API Docs](https://api-docs.deepseek.com/quick_start/pricing/) (primary)
10. [DeepSeek V4 Pro 0813 (max) | Artificial Analysis](https://artificialanalysis.ai/models/deepseek-v4-pro)
11. [www.tbench.ai](https://www.tbench.ai/leaderboard/terminal-bench/2.1?agents=Terminus+2)
12. [Pricing | Claude API Docs](https://platform.claude.com/docs/en/about-claude/pricing) (primary)
13. [Claude Fable 5 overview | Claude API Docs](https://platform.claude.com/docs/en/models/fable-5/overview) (primary)
14. [the-decoder.com](https://the-decoder.com/deepseek-launches-an-improved-v4-pro-model-raises-api-prices-and-makes-its-agent-software-open-source/)
15. [artificialanalysis.ai](https://artificialanalysis.ai/models/releases/comparisons/gpt-6-astra-vs-deepseek-v4-pro)
16. [artificialanalysis.ai](https://artificialanalysis.ai/evaluations/terminalbench-2-1?models=deepseek-v4-pro)
17. [ClinePass | Cline Docs](https://docs.cline.bot/getting-started/clinepass) (primary)
18. [Try DeepSeek V4 Pro in Cline](https://cline.bot/models/deepseek-v4-pro) (primary)
19. [DeepSeek-V4.1-Flash: Smarter, Faster, More Efficient | DeepSeek API Docs](https://api-docs.deepseek.com/news/news260910) (primary)

Author: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Evals](https://newruntime.com/topics/evals.md): Explore the evals topic hub.
- topic hub: [Research](https://newruntime.com/topics/research.md): Explore the research topic hub.
- collection: [Analysis](https://newruntime.com/analysis.json): Open the latest authorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)