Cline posted the chart first: DeepSeek dated the flagship API V4-Pro 0813, and on Terminal-Bench 2.1 it sits at 87.9 — 0.1 behind Fable 5 at 88.0, 2.9 ahead of Opus 4.8, and 15.8 points above the April V4-Pro Preview at 72.1. Same $0.435 / $0.87 per million tokens as the preview. Same 1.6T / 49B active MoE and 1M context we already covered in the April API field note and the May V4-Pro economics piece. This is a checkpoint, not a new family.
The tweet called it "up 15.8%." The arithmetic is 15.8 percentage points (72.1 → 87.9). Relative gain is closer to 22%. The interesting claim is not the percent sign. It is that DeepSeek closed almost the entire Preview-to-Fable gap without raising list price.
TL;DR — what people are actually asking
| Question | Direct answer |
|---|---|
| What shipped? | A dated V4-Pro 0813 checkpoint on the existing deepseek-v4-pro ID |
| When? | Docs/API date 0813 (August 13); Cline's post went up the evening of August 12, 2026 |
| Terminal-Bench 2.1? | 87.9 vs Preview 72.1, Opus 4.8 85.0, Fable 5 88.0, V4-Flash 82.7 |
| Price? | $0.435 in / $0.87 out per 1M tokens (Cline's card; verify live DeepSeek pricing) |
| "57x cheaper"? | Output vs Fable 5's $50/M. Input is ~23x vs Fable's $10/M |
| New architecture? | No. Still 1.6T total / 49B active / 1M context |
| Same as Flash 0731? | No. Flash is the 284B / 13B cheap daily driver; Pro is the flagship |
| Should I switch today? | Point your harness at deepseek-v4-pro and A/B your own terminal tasks |
The chart, with the caveats attached
Cline, August 12, 2026: "Latest reported scores · price per 1M tokens (in / out)." Not an explainx.ai re-run.
| Model | Terminal-Bench 2.1 | List price (in / out per 1M) |
|---|---|---|
| Fable 5 | 88.0 | $10 / $50 |
| V4-Pro 0813 | 87.9 | $0.435 / $0.87 |
| Opus 4.8 | 85.0 | $5 / $25 |
| V4-Flash | 82.7 | $0.14 / $0.28 |
| V4-Pro Preview | 72.1 | $0.435 / $0.87 |
Three things the chart does not say:
- It is Terminal-Bench 2.1, not 3.0. Grok 4.6's official card quotes Terminal-Bench v3.0 (Grok 4.6 High at 26%). Do not rank 87.9 against 26. Different suite, different year of tasks. For what TB 2.x actually measures, start at explainx.ai's Terminal-Bench 2.0 guide.
- "Latest reported scores" means mixed sources. Treat it like every other vendor-adjacent leaderboard: useful for triage, not a purchase order.
- Flash at 82.7 is still the volume play. If your loop is CI fixes and test generation, the Flash 0731 ARC-AGI cost-per-task write-up is the more relevant DeepSeek SKU. Pro 0813 is what you reach for when Flash is not enough.
What "57x cheaper" actually means on a bill
Cline's line — "Fable 5 performance at ~57x cheaper cost" — is output-token list price: $50 ÷ $0.87 ≈ 57.5. Input is $10 ÷ $0.435 ≈ 23. Agent traces are mostly repeated input. If your harness hits DeepSeek cache, the number that moves the invoice is cache-read, not the $0.87 output cell. We walked through that math on Flash in August; the same prompt-caching playbook applies to Pro.
Also: DeepSeek already warned of a "significant" API price increase with no date. 0813 does not reset that notice. Budget the ratio as today's list, not a five-year contract.
Is this a new model or a silent checkpoint?
Checkpoint. The April preview already advertised 1.6T / 49B / 1M. The May builder piece already treated V4-Pro as the agentic flagship next to Flash. Chinese-language coverage of the 0813 date (DeepSeek docs updating the model stamp, no price change) matches Cline: same SKU, better agent evals.
That is the same shipping style as Flash-0731 — a date suffix, not a marketing site. If your client still sends deepseek-v4-pro, you likely already have 0813 on the official API. Confirm the date string in the provider catalog before you celebrate; gateways lag.
Chinese posts circulating overnight also cite jumps on Cybergym, DeepSWE, and AutomationBench. Those are not on Cline's English chart. Until DeepSeek publishes a model card for 0813, treat extra benches as unverified social numbers.
What to do in the harness this week
- If you already call
deepseek-v4-pro: look at traces from August 13 onward. You may already be on 0813. - If you route Flash for everything: keep Flash as the default. Add a Pro 0813 lane for long terminal sessions, multi-file repairs, and anything that failed Flash last month.
- If you are on Fable 5 for volume coding: A/B 50 tasks you actually pay for. A 0.1 TB 2.1 gap does not tell you whether Pro will miss the one production incident Fable would catch.
- Encode the workflow, not the model. The agent skills guide and loop engineering still dominate reliability. A cheaper base model does not fix a sloppy tool schema.
Cline said 0813 is on ClinePass. That is their bundle, not a DeepSeek requirement. Any OpenAI-compatible or Anthropic-compatible client that already talks to DeepSeek can send the same model ID — the pattern from the preview migration note.
Honest limitations
- Primary English source is a coding-agent vendor chart, not a DeepSeek blog with methodology.
- TB 2.1 ≠ TB 3.0. Mixing them with Grok 4.6 is a category error.
- List prices ignore cache, retries, and thinking tokens.
- The price-hike warning is still outstanding.
- 0.1 behind Fable 5 on one bench is not "beats Fable 5." It is "close enough to force a bake-off."
Related on explainx.ai
- DeepSeek V4 preview — API IDs, 1M context, migration
- DeepSeek V4-Pro benchmarks and API economics (May 2026)
- DeepSeek V4-Pro pricing disruption
- DeepSeek-V4-Flash-0731 — Codex, Responses API, $0.14/$0.28
- Flash 0731 on ARC-AGI at $0.02/task
- DeepSeek API price-increase warning
- Terminal-Bench 2.0 — what the suite actually tests
- Grok 4.6 launch evals (Terminal-Bench v3.0, different suite)
- What is an agent harness?
- Prompt caching for LLM cost
- Browse agent skills
Primary sources: Cline post on X (August 12, 2026) with Terminal-Bench 2.1 price/score card · DeepSeek V4-Pro architecture as previously documented (1.6T / 49B / 1M) · DeepSeek Models & Pricing page for live rates
Scores and list prices above reflect Cline's August 12, 2026 chart and prior DeepSeek V4 documentation. They are not an independent re-run by explainx.ai. Confirm the 0813 date on your provider and current per-token rates before you change production routing. Follow @explainx_ai for updates.