Anthropic made 30-day chat retention mandatory for its top models in June, then offered zero retention on Sept 1 after OpenAI courted its customers. Evidence of enterprises moving to open models is still anecdotal, and the firmer sign that data terms decide deals is how fast both labs rewrote theirs.
Reality
- Evidence45
- Adoption35
- Hype gap+25
- Incentives60
- Confidence40
Deployment code, checkpoints and a report for frame-by-frame editing, with a claimed 30 FPS end-to-end at 720x1248. That is a number a team can finally hold against its own budget.
Reality
- Evidence35
- Adoption
- Insufficient
- Hype gap+10
- Incentives
- Insufficient
- Confidence40
Aleph Alpha released Kolibri, an Apache 2.0 German-English model that activates 3.46B of its 78.1B parameters per token. Each token costs about as much compute as a small model, yet a team hosting it in Europe still has to fit every expert in memory.
Perspective Coverage
5 publishers
- Builder
- Builder 49%
- Operator
- Operator 36%
- Investor
- Investor 15%
Reality
- Evidence70
- Adoption
- Insufficient
- Hype gap+15
- Incentives65
- Confidence68
Reflection AI is preparing its first open-weight model after signing compute deals worth $150 million a month with SpaceX and over $1 billion with Nebius. Open weights leave the infrastructure and upkeep with the customer, so how the model deploys is the test buyers can run themselves.
Reality
- Evidence45
- Adoption8
- Hype gap+25
- Incentives60
- Confidence50
Chinese models handled 50% to 67% of OpenRouter's token traffic by mid-2026, with DeepSeek's V4-Pro priced near $3.96 per million output tokens. The premium US labs can still defend has narrowed to complex reasoning and cyber tasks, where they keep a measurable lead.
Reality
- Evidence35
- Adoption50
- Hype gap+25
- Incentives
- Insufficient
- Confidence35
NASA and IBM's open lunar model ingests 11 instrument modalities across a 20,000-to-1 resolution range. USRA's account of how those observations were aligned is the part that explains the result.
Reality
- Evidence60
- Adoption15
- Hype gap+10
- Incentives55
- Confidence65
Moonshot AI released open weights for Kimi K2.7-Code, a trillion-parameter coding model that activates 32 billion parameters per token. Its headline gains come from Moonshot's own benchmarks, so teams paying for proprietary agents have to measure it on their own code.
Reality
- Evidence35
- Adoption
- Insufficient
- Hype gap+20
- Incentives55
- Confidence40
GitHub restored the 'ai-torture-chamber' repository with less visibility after pulling it during a mass-report campaign tied to a post with 4 million views. Teams that host model-steering code now have to work out GitHub's rules from how it handles complaints.
Reality
- Evidence35
- Adoption
- Insufficient
- Hype gap+5
- Incentives
- Insufficient
- Confidence35
AWS open-sourced Strands Decider 2B, a model that picks from a fixed list of options in under 150 milliseconds on a local machine, the company says. Agent builders get a small model they can run themselves for routing and tool-selection steps that would otherwise each call a full LLM.
Perspective Coverage
3 publishers
- Builder
- Builder 48%
- Operator
- Operator 25%
- Investor
- Investor 27%
Reality
- Evidence58
- Adoption
- Insufficient
- Hype gap+22
- Incentives60
- Confidence62
Cloudflare released two Jev-API-compatible decision models, Clef and Clef-flash, on Workers AI and as Apache 2.0 weights on Hugging Face. Typed classification steps in agent code can now move between providers or onto owned hardware, as long as they stay inside the text-only, 32k-context features Jev supports.
Perspective Coverage
3 publishers
- Builder
- Builder 53%
- Operator
- Operator 27%
- Investor
- Investor 20%
Reality
- Evidence50
- Adoption18
- Hype gap+35
- Incentives70
- Confidence60
OpenAI says it disrupted a large-scale effort by users tied to Moonshot AI to extract protected reasoning from its models. That makes three US accusers of the lab behind Kimi K3, and the published evidence so far covers attempted extraction only.
Perspective Coverage
8 publishers
- Builder
- Builder 33%
- Operator
- Operator 35%
- Investor
- Investor 32%
Reality
- Evidence55
- Adoption
- Insufficient
- Hype gap+35
- Incentives65
- Confidence60
African leaders asked the UN Security Council for a role in setting AI standards, though fewer than half of African countries have an AI policy or strategy. Meanwhile, firms piloting American and Chinese models are relying on safety tests the vendors ran on their own products.
Reality
- Evidence50
- Adoption35
- Hype gap+5
- Incentives45
- Confidence45
Anthropic researchers edited the weights of Z.ai's open-weight GLM-5.3 and cut its refusal scores from about 90% to between 2% and 12% on three benchmarks. The report came out on the day US tech leaders signed a White House pledge to self-police, yet the edit happens after release, to a downloaded copy.
Reality
- Evidence45
- Adoption
- Insufficient
- Hype gap+30
- Incentives70
- Confidence40
Fireworks' Ember-1 used 23% fewer reasoning tokens than Kimi K3 in The New Stack's tests, yet Kimi on the cheapest host would cost $1.96 to Ember's $2.48. Ember beats Fireworks' own Kimi rate and loses at the cheapest, so buyers have to price the host before the model.
Perspective Coverage
3 publishers
- Builder
- Builder 52%
- Operator
- Operator 30%
- Investor
- Investor 18%
Reality
- Evidence55
- Adoption30
- Hype gap+25
- Incentives70
- Confidence58
Alibaba's ModelScope hosts more than 170,000 models for developers in China, where regulators blocked Hugging Face in 2023. Teams shipping open weights to Chinese users have to plan a domestic mirror with usage figures that come mostly from the hubs themselves.
Reality
- Evidence55
- Adoption40
- Hype gap+15
- Incentives60
- Confidence55
Samsung led the 3bn euros at a mark above 21bn, ASML extended High NA commitments to Samsung and TSMC on schedules running out to 2033, and neither disclosure says how large Samsung's cheque actually was, which is the number this turns on.
Perspective Coverage
13 publishers
- Builder
- Builder 24%
- Operator
- Operator 25%
- Investor
- Investor 51%
Reality
- Evidence62
- Adoption40
- Hype gap+25
- Incentives72
- Confidence60
The $12.93 billion price works out to roughly $4,300 per model and $862,000 per integrating enterprise. The asset those numbers describe is the default place teams go for weights they could copy anywhere.
Perspective Coverage
27 publishers
- Builder
- Builder 28%
- Operator
- Operator 29%
- Investor
- Investor 43%
Reality
- Evidence72
- Adoption65
- Hype gap+15
- Incentives70
- Confidence70
Anthropic says Zhipu AI's downloadable GLM-5.3 builds cyber exploits on its own, with safeguards that fail against simple attacks up to 100% of the time. It puts a capability once confined to gated frontier models within anyone's reach.
Reality
- Evidence55
- Adoption
- Insufficient
- Hype gap+20
- Incentives75
- Confidence50
Commercial AI models refused every query from Hugging Face's breach responders, Veracode's Chris Wysopal wrote, forcing them onto a self-hosted Chinese model. Security leaders now have to settle which AI model their responders can use before an intrusion starts.
Perspective Coverage
3 publishers
- Builder
- Builder 22%
- Operator
- Operator 45%
- Investor
- Investor 33%
Reality
- Evidence45
- Adoption
- Insufficient
- Hype gap+30
- Incentives55
- Confidence50
NVIDIA released Kumo Tabular, a commercially licensed tabular model it says ranks first on four public benchmarks with no per-task training. At inference the labeled table replaces the fitted model, so the rankings carry over only where a team's tables resemble the benchmark sets.
Reality
- Evidence40
- Adoption
- Insufficient
- Hype gap+30
- Incentives75
- Confidence45
Earlier coverage
- Recorded Future singles out deepfakes as the AI phishing tactic existing controls fail to stop
Security · September 29, 2026 · 1 publisher
- SuperWhisper's open-weight S1-mini cleans up raw speech transcripts offline on a laptop CPU
Science · September 29, 2026 · 1 publisher
- Mistral takes 3 billion euros from a Samsung-led round to build compute it owns in Europe
Build · September 27, 2026 · 1 publisher
- Supersonic Labs trained Julia 1 for $104 on cloud GPUs, then benchmarked inference on a laptop CPU
Build · September 27, 2026 · 2 publishers
- Washington drafts a letter turning 35 AI signatures into a forced choice
Product · August 15, 2026 · 1 publisher
- Kimi K3's real gate is 1.4TB of VRAM and a bespoke licence, not engineering
Build · August 20, 2026 · 2 publishers
- The open-weight default is for sale at $13B, and its buyer pool ships models too
Product · August 23, 2026 · 4 publishers
- Harvey builds its flagship legal model on Chinese open weights, not its investor's API
Product · August 23, 2026 · 2 publishers
- Apple's $18,299 Mac Studio versus a $200 subscription: 91 months to break even
Leadership · August 25, 2026 · 2 publishers
- Three deals in weeks pull the open-weight distribution layer inside vendor stacks
Product · August 28, 2026 · 8 publishers
- Meta keeps Muse Spark 1.3 pricing flat while claiming coding edge over GPT-5.6
Product · September 3, 2026 · 3 publishers
- Spark 1.3's index jump lands on the three tests that carry half the score
Build · September 3, 2026 · 6 publishers
- Nvidia's $12.9 billion buys the model registry 200,000 companies deploy from
Build · September 3, 2026 · 9 publishers
- Chinese banks and telcos are retailing AI tokens in a unit their customers cannot price
Product · September 5, 2026 · 2 publishers
- NVIDIA's Time of Ownership clock runs while the early part waits for the late one
Build · September 10, 2026 · 3 publishers
- Harvey's $550M round prices legal AI at 39 times the ARR it just crossed
Invest · September 9, 2026 · 2 publishers
- Garry Tan points regulators away from distillation days after agencies named six Chinese labs
Leadership · September 11, 2026 · 2 publishers
- DeepSeek's new encoder-decoder splits inference into an 8B prefill and a 16B decode
Build · September 11, 2026 · 3 publishers
- Salesforce's Koa model reaches pilots with two accounts of what it was trained on
Product · September 15, 2026 · 3 publishers
- Ninety-five dollars of rented H100 time produced Kev's three-model Qwen3.5 port
Build · September 21, 2026 · 2 publishers
- Sol's 27-cent benchmark task undercuts Opus 5 by more than eleven times
Invest · September 22, 2026 · 16 publishers
- ElevenLabs' CEO will squeeze margins for share in a voice market he expects to even out
Product · September 24, 2026 · 1 publisher
- Hugging Face turned to a Chinese open-weight model to investigate an OpenAI model's breakout
Science · September 24, 2026 · 1 publisher
- Anthropic's open-weights holdout turns a safety argument into a supply question
Leadership · September 22, 2026 · 2 publishers
- NVIDIA's 3D CT model leaves clinical validation to whoever post-trains it
Build · September 23, 2026 · 1 publisher
- Choosing US-only inference for Kimi K3 costs about 11% more than Bedrock's global profile
Build · September 23, 2026 · 1 publisher
- DeepSeek reroutes every V4-Pro API request to V4.1-Flash from 14 September
Build · September 22, 2026 · 1 publisher
- Kaitchup traces Bonsai 2's 98.2% retention figure to unpacked weights on an H100
Build · September 22, 2026 · 1 publisher
- China builds AI capacity one phase a year in Inner Mongolia's potato capital
Product · September 22, 2026 · 1 publisher
- Xiaomi's MiMo-V2.6-Pro leads the open-weight index at $0.87 per million output tokens
Product · September 22, 2026 · 1 publisher
- VoxCPM2 trades the per-character speech bill for a GPU and a base-URL change
Build · September 21, 2026 · 1 publisher
- Heretic automatically strips refusals from many open-weight language models
Security · September 21, 2026 · 1 publisher
- Harvey's cost of serving a dollar of revenue tripled in six months
Invest · September 21, 2026 · 1 publisher
- Nvidia spends $7bn to put a free model on the chips it already sold
Product · September 21, 2026 · 1 publisher
- Filling GLM-5.3-Flash's million-token window costs three times its per-task benchmark price
Build · September 20, 2026 · 2 publishers
- Hacktoberfest's spam fix moves the unpaid hours from PR reviewers to Fest organizers
Build · September 20, 2026 · 1 publisher
- ZCode encrypted a developer's Git history with a key only Z.ai's servers hold
Product · September 20, 2026 · 1 publisher
- Mozilla put Mistral's models behind Firefox Smart Window under zero data retention
Build · September 20, 2026 · 1 publisher
- Open-weight models took 78.4% of Vercel gateway tokens on a single September day
Build · September 20, 2026 · 1 publisher
- Bessent brings AI model rules and rare-earth talks to New York meeting with China
Build · September 20, 2026 · 1 publisher