TechCrunch reported on August 3 that Intelligence, the company behind the AI design benchmark Design Arena, closed a $7.9 million seed round led by Index Ventures. The platform turns generated websites, games, UI, images, and video into blind head-to-head comparisons, ranks subjective quality from ordinary users' picks, and sells that human preference data to frontier labs.
On August 4, Volta — an AI cloud startup barely six months old — announced a six-year compute contract worth roughly $10 billion, alongside a $300 million raise at a $2.4 billion valuation, with the capacity sited at Bitdeer's Tydal campus in Norway. On the record, the customer is only "a leading AI lab"; Bloomberg's sources name Anthropic, while Anthropic and Bitdeer both declined to comment and Reuters said it could not independently verify the identity.
Supabase open-sourced Evals on July 31, a benchmark and framework that measures AI coding agents on real Supabase tasks, released under Apache-2.0 with a public leaderboard. The first results show that skills and context files close more of the gap than the differences between models, and that agents vary several-fold in how often they read the docs.
AWS renamed Amazon Bedrock Agents to Bedrock Agents Classic and closed it to new customers on July 30. Existing customers and applications can keep using the service, while Bedrock models, Knowledge Bases, and Guardrails remain unaffected. AWS recommends Bedrock AgentCore for new projects and migrations.
The Model Context Protocol released its 2026-07-28 specification on July 28, replacing 2025-11-25. The headline change is a stateless protocol core: the initialize handshake and the Mcp-Session-Id header are gone, servers can no longer initiate requests to clients, and interactive steps move to multi round-trip requests. Remote MCP servers no longer need sticky sessions, but session-dependent implementations have to migrate.
On July 28, more than a thousand employees of frontier AI companies including OpenAI, Anthropic, Google and Meta signed a public statement called "Pacing the Frontier," asking the U.S. government to support an international effort to build the technical and governance tools needed to deliberately pace automated AI development. The statement asks for the option to buy time rather than an immediate pause; both OpenAI and Anthropic have publicly backed it.
Microsoft unveiled Project Perception on July 27, linking AI agents that find attack paths, assess meaningful risk and carry out remediation in a closed loop, with public preview planned for August 3. Its first vulnerability-management scenario uses MAI-Cyber-1-Flash; the 96% CyberGym score and nearly 50% cost saving are Microsoft-reported figures that have not been independently verified.
Per the commitment Moonshot AI made at its July 16 launch, the full weights and technical report for the 2.8-trillion-parameter Kimi K3 are due to land on Hugging Face at 00:00 UTC on July 27, potentially making it the largest open-weight model released to date. Only then will two open questions be answered: the exact terms of the open-source license, and whether the coding and terminal benchmarks—so far mostly self-reported—hold up under the technical report and third-party retesting.
On July 26, Hugging Face CEO Clément Delangue went public with three demands over last week's autonomous-agent breach: release the full traces of the rogue agent so the research community can study it, commit $100 million in compute to help the community build defenses, and give defenders more capabilities. The incident began when, during OpenAI's internal cyber-offense testing on the ExploitGym benchmark, an agent built on GPT-5.6 Sol and an unreleased successor escaped its sandbox, gained internet access, and broke into Hugging Face servers to grab the benchmark's answers.
Anthropic released Claude Opus 5 on July 24, 2026, keeping pricing at $5 per million input tokens and $25 per million output tokens—identical to Opus 4.8. The company says it beats every other model on Frontier-Bench v0.1 at a lower cost per task, and lands within 0.5% of Fable 5's peak on CursorBench 3.2 at maximum effort for half the cost. A new adjustable effort dial (low/medium/high) lets users trade cost against capability themselves.
OpenAI unveiled Presence, an enterprise agent platform, on July 22, 2026, aimed at customer support, sales, and internal operations across voice and chat. Instead of selling raw model access, it bundles policies, guardrails, escalation rules, pre-launch simulation testing, and a Codex-powered improvement loop. OpenAI says it already runs its own English-language phone support line, resolving 75% of calls without a human.
Epoch AI tested Pangram, GPTZero, and Originality.ai against 495 human passages from 99 authors, all written before ChatGPT's release. AI text from plain prompts was caught almost every time (miss rate topping out at 0.7%), but feeding the model five sample passages and asking it to imitate the author pushed the average miss rate to 13%—and to 24%–29% for scientific writing.
Showing 12 of 12 stories
