Analysis
The Release
DeepSeek pushed DeepSeek-V4-Pro-0813 to general availability, updating its official API pricing page on August 12 to reflect the new build, according to Tech Times and corroborated by Yahoo. The GA release focuses on agentic capability -- tasks where a model uses tools, writes and executes code, and completes multi-step workflows without a human stepping in at each turn, rather than chasing a higher score on general chat benchmarks.
The Numbers DeepSeek Is Showing
DeepSeek's self-reported benchmarks show some of the largest single-release jumps of any model this year:
- Terminal Bench 2.1 -- 87.9, up from 72.1 in the preview build
- DeepSWE -- 62.7, up from 12.8
- CyberGym -- 83.3, up from 52.7
- NL2Repo -- 61.5
Those are gains of up to 49.9 percentage points versus the preview. The model supports a context window up to 1 million tokens, can generate outputs as long as 384,000 tokens, and lets developers toggle between a thinking mode and a faster non-thinking mode depending on the task.
Company Background
DeepSeek was founded in 2023, spun out of the Chinese quantitative hedge fund High-Flyer, and built its early reputation on delivering frontier-adjacent capability at a fraction of the compute cost US labs assumed was necessary. Pulse covered DeepSeek's V4-Flash costing roughly 3 cents per benchmark test to run -- more than 100 times cheaper than Anthropic's Claude Fable 5, albeit at meaningfully lower intelligence scores.
The funding trajectory has moved just as fast as the model releases:
- June 2026 -- closed a $7.4 billion round backed by Tencent, CATL, JD.com, NetEase, and IDG Capital at a valuation north of $50 billion
- July 2026 -- moved into fresh talks for a raise at roughly $71 billion, Pulse reported
- This month -- Pulse most recently reported DeepSeek resumed a round targeting about 50 billion yuan at a pre-money valuation near 500 billion yuan, expected to close late August -- the same window this release lands in
The Competitive Field
V4 Pro's agentic focus puts it in direct competition with OpenAI's GPT-5.6 family, Anthropic's Claude Code line, and Google's Gemini 3.6 Flash on tool-use and coding-agent benchmarks -- the same category Cognition's Devin and Cursor-maker Anysphere (now part of SpaceX) compete in as products rather than base models. Domestically, DeepSeek is racing Alibaba's Qwen and Moonshot AI's Kimi, both of which Pulse has covered as also preparing IPO paths this year.
The Counterweight
Every one of the headline benchmark gains is vendor-reported, and that's a real limitation worth noting before treating the numbers as settled. No independent evaluator -- not Artificial Analysis, not a third-party lab -- has yet replicated DeepSeek's Terminal Bench or DeepSWE numbers for this specific build, and a 49.9-percentage-point jump on DeepSWE in particular is large enough that it deserves outside verification before it's treated as fact rather than marketing. Releasing a stronger model just four days before a price increase also isn't neutral timing: it gives developers a reason to lock in usage at current rates before the August 16 hike, which is a demand-pull tactic as much as a pure capability story.
Watch whether independent benchmarks confirm the jump once V4 Pro sees wider use, and whether the August 16 price increase changes DeepSeek's cost advantage over Western frontier models now that the gap has narrowed.
