DeepSeek ships V4 Pro to general availability with zero independently verified benchmark scores
The build's version string surfaced first in DeepSeek's own price list, with no separate release announcement.
DeepSeek ships V4 Pro to general availability with zero independently verified benchmark scores1
DeepSeek moved V4-Pro-0813 from preview to general availability on its app, web console, and API in a roughly 32-hour window around August 12 and 13.12 The model card lists 1.6 trillion total parameters with 49 billion active per token in its mixture-of-experts routing, a 1,000,000-token context window, and MIT-licensed open weights.3
| Benchmark | Score |
|---|---|
| SWE-bench Verified | 80.6% |
| Terminal-Bench 2.0 | 67.9% |
| GPQA Diamond | 90.1% |
| MMLU-Pro | 87.5% |
| LiveCodeBench | 93.5% |
| Codeforces rating | 3,206 |
DeepSeek's own changelog documents the GA transition only as a dated entry.1 The most detailed independent read of the release found that the version string surfaced first in DeepSeek's price list, with no separate announcement, and that accounts still calling the model by its unversioned alias were upgraded automatically, without a version pin.2 Of the benchmark table's numbers, it states plainly: "At the time of writing, no independent reproduction of any GA-specific figure exists — normal for a release this fresh, and still the single most important caveat on the table."2 Pricing changes on August 16: peak-hour output rises from a flat $0.87 per million tokens to $3.96, with off-peak rates held to half that.1
IBM joins OpenAI's Elite partner tier without disclosing deal terms4
IBM and OpenAI announced a partnership on August 13 that embeds GPT-5.6 and OpenAI's Codex and ChatGPT Work products inside IBM Consulting Advantage.4 The deal is backed by a dedicated "OpenAI Practice" IBM describes as staffed by "thousands of consultants and engineers" holding OpenAI Partner Network certifications, and it makes IBM a member of OpenAI's "Elite" partner tier.4 Initial focus falls on legacy-system modernization, software development, and cybersecurity through OpenAI's Daybreak Cyber Partner Program alongside IBM's own Autonomous Security offering, aimed first at financial-services, government, telecom, and retail clients.4 Neither company's release states a dollar figure, a consultant headcount, or a contract length, and TechCrunch's independent read of the announcement found none disclosed to any outlet.45 The deal lands weeks after IBM cut its own revenue forecast on weaker-than-expected quarterly results, and follows comparable enterprise partnerships OpenAI has struck with Infosys and Tata Consultancy Services.5
Gemini 3.7 Flash trails three rivals on an independent benchmark Google's own testing skipped6
Gemini 3.7 Flash, which Google released August 13 at half its eventual price through the end of the year, now has an independent measurement that runs outside Google's own launch benchmarks.6 On Artificial Analysis's GDPVal-AA v2 leaderboard, the model scores 1,525, a 103-point gain over Gemini 3.6 Flash.7 That still trails three named rivals on the same scale: Meta's Muse Spark 1.2 at 1,628, Anthropic's Claude Sonnet 5 at 1,598, and OpenAI's GPT-5.6 Terra at 1,578.7 Google's launch benchmarks and Artificial Analysis's leaderboards don't share a single test, so the two scores measure different things rather than confirming or disputing each other.678 On Artificial Analysis's separate Intelligence Index, a nine-evaluation composite, the model scores 56, ranking 17th of the 188 models the organization has tested and landing above the 34-point median for its price tier.8
Atlantic overturning circulation's collapse threshold swings from 2°C to 5.5°C depending on warming speed9
The Atlantic Meridional Overturning Circulation (AMOC), the current system that carries heat northward through the Atlantic, has been described as having one collapse threshold. That estimate is +4.0°C of global warming, with an uncertainty range of 1.4°C to 8°C.9 A Nature Climate Change paper published August 13 by Utrecht University's René M. van Westen, Reyk Börner, and Henk A. Dijkstra ties that threshold to warming speed instead of warming's eventual peak.9 In a climate-model simulation, a slow CO2 ramp of 0.5 parts per million a year keeps the circulation stable to +5.5°C of warming. An otherwise identical fast ramp collapses it at +2.0°C instead.9 Slow forcing lets the ocean's surface and interior properties adjust together, aided by more evaporation and less sea ice, offsetting the feedbacks that would otherwise destabilize the current.9 "Our results show there is not necessarily a fixed temperature beyond which the AMOC inevitably collapses," van Westen said. "The stability of the circulation depends on how fast the climate is changing."10 Utrecht University's own summary puts the critical warming rate at about 0.3°C per decade, a pace it says current emissions trends are already approaching — though the figure does not appear in the paper's own public abstract.10
Sources
- DeepSeek · API changelog: DeepSeek-V4-Pro-0813 general availability
- Digital Applied · DeepSeek V4 Pro GA: the official release
- Unite.AI · DeepSeek ships V4 Pro as its flagship model leaves preview
- IBM · IBM partners with OpenAI to accelerate secure AI deployment for enterprises across core operations
- TechCrunch · IBM partners with OpenAI to bolster enterprise AI push
- Google · Introducing Gemini 3.7 Flash
- Artificial Analysis · GDPVal-AA v2 leaderboard
- Artificial Analysis · Gemini 3.7 Flash model page
- Nature Climate Change · van Westen, Börner & Dijkstra, failure to track a stable AMOC state under rapid climate change
- Phys.org · Utrecht University release on AMOC tipping-rate paper