AI-news notes.

Jeff Bezos's physical-AI lab closes a $10B round at a $38B valuation, led by BlackRock and JPMorgan. Prometheus builds AI systems that learn through real-world interaction and physics rather than text or images — a deliberate contrast to the web-trained foundation model stack. Bezos is reportedly structuring a broader industrial-AI holding company targeting up to $100B. Physical AI is the next frontier-model arms race, and Bezos is running the first lap.

Cognition — maker of Devin, the first broadly-deployed AI software engineer — enters funding talks at a $25B valuation, more than double its September 2025 mark. Devin's ARR grew from $1M in September 2024 to $73M by June 2025, a 73x run-rate in nine months. The round is not closed and terms may shift, but the trajectory is the point: agentic coding products are compounding at a pace that makes 2024 benchmarks look quaint.

Anthropic launches a cybersecurity coalition with AWS, Apple, Google, Microsoft, and others — backed by Claude Mythos Preview, a new frontier model optimized for finding software vulnerabilities. The initiative commits $100M in model credits and $4M to open-source security. The framing: ensure advanced AI cyber capabilities reach defenders before attackers. The notable subtext: Claude Mythos is the first public signal of a post-Opus model in the pipeline.

Google DeepMind ships Gemma 4 in four sizes — E2B, E4B, 26B MoE, and 31B Dense — distilled from Gemini 3 and released under Apache 2.0 for the first time in the family's history. The 31B scores 89.2% on AIME 2026 (+68 points over Gemma 3), 80% on LiveCodeBench, and 86.4% on agentic benchmarks. It's the #3 open model on the Arena leaderboard — a 31B model outperforming 400B-class competitors. The open-weight frontier just moved again.

OpenAI closes the largest private funding round in history: $122B with Amazon ($50B), Nvidia ($30B), and SoftBank ($30B) as lead investors. Valuation hits $852B. For context: that's larger than most sovereign wealth funds and nearly every public tech company outside the Mag-7. The capital is explicitly for compute and infrastructure, not product. The race for GPU clusters is now denominated in hundreds of billions.

OpenAI deprecates GPT-5.1 and ships GPT-5.4, GPT-5.4 Thinking, and GPT-5.4 mini. The model line restructure collapses the older GPT-5.x variants into three tiers: instant, thinking, and pro — mirroring how Anthropic structured the Claude 4 family. ChatGPT also gains CarPlay integration, a File Library, and interactive math and science modules for 70+ topics. The platform is quietly becoming an OS-level interface.

Meta's first major model release since acquiring Scale AI's Alexandr Wang scores #4 on the Artificial Analysis Intelligence Index with strong multimodal, reasoning, health, and agentic results — at a fraction of the compute cost of Llama 4 Maverick. Meta guided $115–135B in AI capex for 2026. The signal: the open-model era is over; Meta is building proprietary closed frontier models now.

xAI ships Grok 4.20 with the strongest current-events accuracy of any frontier model at release — a direct product of real-time X social data integration that no other lab has access to at equivalent scale. The model closes the factuality gap that plagued Grok 3, with Grok 4.20 leading on news accuracy within a 30-day window. The 4-agent architecture ships here: Grok as coordinator, Harper as research, Benjamin for logic and mathematics, and Lucas as contrarian analysis — all running in parallel and cross-verifying outputs before a response surfaces. Intelligence Index score: 49, placing it competitively with but not clearly ahead of GPT-5.4 or Claude Opus 4.7 on general benchmarks. Simultaneously, xAI closes a $20B Series E with Nvidia, Cisco, QIA, and others. The pairing is intentional: the money goes straight to compute, the model is the proof of concept that it's being spent well.

Anthropic spins out a dedicated research organization led by co-founder Jack Clark — who spent five years at OpenAI as Policy Director before leaving to co-found Anthropic, and who runs Import AI, a newsletter with 70,000 weekly subscribers that has tracked every major model release since 2017. The Institute consolidates three teams: Frontier Red Team (stress-testing models at the outer edge of their capabilities), Societal Impacts (tracking real-world deployment effects), and Economic Research (measuring labor-market shifts as AI scales). The mandate is deliberately empirical: produce primary data about how AI is affecting workers and economies, rather than relying on modeling or extrapolation, and publish findings even when they're uncomfortable. Hires include Matt Botvinick from Google DeepMind on AI and rule of law, Anton Korinek from UVA on economic transformation, and Zoë Hitzig, formerly of OpenAI, bridging economics and model development. The structural move matters: by separating safety and societal research from the product org, Anthropic is betting that credibility requires independence — a research institute that reports to the same team shipping Claude has a conflict of interest on every finding.

Karpathy releases a 630-line open-source script that lets an AI agent autonomously run ML experiments on a fixed compute budget — hypothesize, modify code, run, collect results, repeat overnight. After two days on a GPU, the agent found ~700 improvements, with ~20 transferring to larger models and delivering an 11% efficiency gain (Time-to-GPT-2 from 2.02h → 1.80h). The repo hit 85K+ stars. The cleanest possible signal that AI agents are now the primary tool for research.

Within days of Karpathy's release, Tobi Lütke adapts autoresearch for a Shopify model training run — a 0.8B parameter model with a Raspberry Pi-based compute loop — and reports a 19% validation improvement. The public adaptation loop between practitioner and researcher signaled a new kind of open science: fork, run overnight, post results by morning.

Anthropic's flagship at release, later succeeded by Opus 4.7. The headline capability is multi-agent coordination: Opus 4.6 acts as an orchestrator that spins up and manages a team of specialized subagents, each running tool calls and accumulating context independently, then consolidates their outputs — distinct from earlier single-agent loops that just re-called the same model. The 128K max output token limit makes full codebase rewrites and long-form document generation tractable in a single pass; prior models capped at 8K. Extended thinking with tool use lets Claude alternate between chain-of-thought reasoning and live tool calls mid-problem, rather than reasoning first and then executing separately. On OSWorld — a benchmark that measures autonomous computer use across real GUI environments — it hits 72.7%, the state of the art at release; context is that Claude 3.5 Sonnet scored around 22% when computer use first shipped in late 2024, so the jump in eighteen months is nearly 50 percentage points. SWE-bench Verified: 80.8%, just above Sonnet 4.6 at 79.6%. Priced at $5/$25 per million tokens on a 200K context window.

The first Sonnet-class model to hit 1M token context, which matters because previous Sonnet models topped out at 200K — the jump makes full-repo analysis and large document corpora practical at mid-tier pricing. Ships February 17, 2026. SWE-bench Verified: 79.6%, within 1.2 points of Opus 4.6 (80.8%) at one-third the cost — at release, GPT-4.1 sat around 54% on the same benchmark, so Sonnet 4.6 lands well ahead of the prior OpenAI mid-tier. OSWorld: 72.5%, essentially tied with Opus 4.6 at 72.7%. Terminal-Bench 2.0: 59.1%. Developer telemetry shows users choosing Sonnet 4.6 over Opus 4.5 59% of the time — the preference flip confirms that the performance-per-dollar gap had closed enough that latency and price became the deciding factors. What changed from 4.5: adaptive thinking (the model dynamically scales reasoning depth per task rather than requiring a mode switch), native 1M context, and context compaction for sessions that would otherwise overflow. $3/$15 per million tokens.

Released February 27, 2025, GPT-4.5 is OpenAI's largest non-reasoning model — the explicit bet that raw scale and better training data produce a qualitatively different conversational experience, not just higher benchmark numbers. SimpleQA accuracy: 62.5%, against 47.0% for GPT-4o — it hallucinated 37.1% of the time versus 59.8% for 4o, which is the headline improvement in practice. MMLU: 85.1%, above o3-mini's 81.1% on undergraduate-level knowledge. On math and hard science, o3-mini still wins by significant margins — the tradeoff is explicit. Where 4.5 separates itself is emotional register: it detects sentiment, adjusts tone, and tracks implicit intent across long conversations in ways that prior models treated as secondary concerns. Context window: 128K tokens. Pricing at launch was $75/$150 per million tokens — 5× the cost of GPT-4o — which made it effectively a research and power-user product rather than a default API choice. The positioning is 'know what you mean,' not 'think harder'; it was the last model OpenAI would release before the o-series and GPT-5 line fully absorbed the product roadmap.

Open-sourced March 4, 2026 under Apache 2.0 at github.com/openai/symphony, Symphony is a Codex App Server orchestration spec that turns a Linear board into a persistent agent control plane. The mechanism: every open issue gets a dedicated Codex agent in an isolated workspace; Symphony polls the board continuously, restarts stalled agents, picks up new issues as they appear, and runs the full ticket lifecycle — triage, code, tests, PR — without a human in the loop until review. The reference implementation was written in Elixir by Codex itself in a single pass; OpenAI then had Codex reimplement it in TypeScript, Go, Rust, Java, and Python to stress-test the spec for ambiguities. Internal teams using Symphony reported a 500% increase in landed PRs in the first three weeks — the metric that matters, since merged code is the unit of delivery. The model running each agent is Codex (GPT-5.x family). What makes it production-grade is the failure recovery: agents that crash or stall are automatically restarted, so the board drains rather than stalls.

Huntley's thesis: stop building brick by brick, start programming the loop. A single monolithic, autonomous process that runs continuously — the agent *is* the system. The post reframes agentic development from task execution to loop design. Where your loop is slow or stuck, that's where your growth is.

Anthropic ships Haiku 4.5 on October 15 — the fastest and cheapest model in the Claude 4 family, and the first Haiku-tier model to support both extended thinking and computer use. The benchmark number that reframes what the tier means: 73.3% on SWE-bench Verified — matching Sonnet 4.6 and roughly tied with GPT-5 on the same benchmark, at $1/$5 per million tokens (one-third the Sonnet 4.6 price). Inference speed: more than 2× faster than Sonnet 4.6, ~300ms to first token, 226 characters/second throughput via Vertex AI. Context window: 200K tokens, max output 64K. API model ID: claude-haiku-4-5-20251001. The extended thinking support changes the tier's role structurally: prior Haiku generations were fast completions models, not reasoning models. Haiku 4.5 can allocate a 128K thinking budget for hard sub-problems, which makes it a capable reasoning sub-agent in pipelines where Sonnet or Opus pricing would be cost-prohibitive at volume. The intended deployment pattern Anthropic describes explicitly: Sonnet or Opus orchestrates, multiple Haiku instances execute sub-tasks in parallel. Computer use at Haiku pricing changes the unit economics of GUI automation pipelines that had previously required Sonnet-tier spend. The model ships as the default free-tier model on claude.ai, meaning Haiku 4.5 is the first contact point for every new Claude user — which is a different kind of capability unlock than any enterprise API number.

OpenAI takes the Realtime API from beta to general availability on August 28, paired with gpt-realtime — its most capable speech-to-speech model and a direct replacement for the stitched-together STT/LLM/TTS pipelines that had defined voice AI for three years. The capability jump is measurable: 66.5% on ComplexFuncBench audio eval versus 49.7% for the December 2024 model, meaning the model now reliably calls tools during voice sessions rather than fumbling them. Three new features ship at GA: image input (the model can see what the user is looking at during a voice call), remote MCP server support (live tool calls to external services mid-conversation), and SIP phone integration (connect directly to the public phone network and PBX systems without a telephony middleware layer). Two new voices — Cedar and Marin — are exclusive to the Realtime API. The SIP support is the underappreciated detail: it means the API can replace an IVR system, not just augment a chatbot. The pattern is the same one that played out in text APIs two years earlier — proprietary bespoke pipeline builds become commoditized surface, and the product race moves one level up.

Mistral ships Medium 3.1 on August 12 — a multimodal proprietary model with a custom-trained vision encoder, 128K context, and output at 113 tokens per second, priced at $0.40/$2.00 per million tokens. The pitch is explicit: roughly 90% of Claude Sonnet 3.7's performance at a fraction of the cost, with 74.4% on MMLU Pro and strong multimodal reasoning. The model is fully closed — no weights — which marks a quiet but significant strategic line for Mistral, a company that built its brand on open-weight releases. The commercial logic is clear: open models serve as marketing and research vehicles; Medium 3.1 is the product. For developers who need cost-efficient multimodal inference at production scale, it fills the gap between cheap general models and Claude/GPT tier pricing. The European differentiation adds real purchasing leverage — GDPR compliance, EU data residency, no American hyperscaler dependency in the call path. Mistral is quietly becoming a closed-model company while keeping the open-weight brand active, a dual strategy that lets it compete on both axes simultaneously.

OpenAI ships GPT-5 on August 7 — not as a single new model but as a unified system: a fast tier for most requests, a thinking tier (GPT-5 thinking) for hard problems, and a real-time router that picks between them without user intervention. Benchmarks at release: 74.9% on SWE-bench Verified (leading all models), 94.6% on AIME 2025, 88.4% on GPQA expert science. Context is 400K tokens with up to 128K output. Pricing spans three tiers — standard at $1.25/$10, mini at $0.50/$5, nano at $0.15/$1.50 per million tokens — the first time OpenAI had occupied every price point in a single release family. The architectural story matters more than any individual benchmark: GPT-5 eliminates the manual model selection problem that had accumulated since o1. Developers no longer decide which reasoning class to route a request to; the system decides. The nano tier at $0.15/$1.50 is a pricing signal as much as a product — OpenAI is deliberately preventing smaller providers from owning the cost-sensitive production tier. GPT-5 is the first OpenAI release designed to retire its entire prior generation in one move rather than layering on top.