AI Digest July 31: OpenAI slashes prices, robots get one brain

OpenAI cut GPT-5.6 Luna prices by 80% and moved code auto-review to a ~10× cheaper model, Thinking Machines released the open multimodal Inkling-Small (276B parameters, 12B active), Google unveiled Gemini Robotics 2 — "one brain for any robot" — and cloud agents now write 56% of merged PRs at Cursor.

Today's top stories

1. OpenAI slashes GPT-5.6 prices: Luna drops 80%

OpenAI announced aggressive price cuts: GPT-5.6 Luna is down 80% and Terra down 20%. Alongside them comes Sol Fast — up to 2.5× lower latency at twice the standard price, with "no change in intelligence." The bigger deal for agent workflows: auto-review in the ChatGPT app and Codex CLI is moving from GPT-5.4 to Luna, which OpenAI expects to be roughly 10× cheaper. The company ties the cuts to efficiency work across the whole "model + inference stack + agentic harness" system, and Sam Altman frames it as a shift in the price/performance frontier.

2. Thinking Machines ships Inkling-Small — an open multimodal MoE with 276B parameters

Thinking Machines released Inkling-Small: open weights, natively multimodal, 276B total parameters with just 12B active — roughly a quarter the size of the original Inkling at comparable quality. The model processes audio and images jointly with text and can inspect images with Python code while reasoning. The open stack picked it up on day one: vLLM announced day-0 support, Modal showed single-B300 deployment, and Unsloth published a local-running/GGUF guide. Artificial Analysis scored it 40 on its Intelligence Index — within a point of the flagship Inkling, with particular strength in coding and science.

3. Gemini Robotics 2: "one brain for any robot"

Google DeepMind launched Gemini Robotics 2 — a stack where a single checkpoint controls different robot types: whole-body humanoid control, fine dexterity (knot-tying, screwing in a bulb), and multi-robot collaboration. It ships with Gemini Robotics ER 2, a high-level embodied reasoning model that observes, plans, tracks progress and recovers from failed steps during multi-minute tasks. A practical detail: On-Device 2 reportedly adapts to a new two-arm robot with fewer than 200 examples. ER 2 is already available via API, and NVIDIA promptly pitched Jetson AGX Thor as the local hardware for humanoids.

4. Cursor: cloud agents now write 56% of merged PRs

The strongest production-agents datapoint of the day comes from Cursor: in December, cloud agents produced 1 in 10 merged PRs; now it's 56%. The company attributes the jump to giving agents their own cloud computers and letting them improve their environments over time. The same trend elsewhere: Cognition added native GitHub stacked PR support for agents, Devin got cloud macOS machines with Xcode and simulator access to build native iOS apps, and a Cognition engineer admitted he never set up a local laptop after joining — all his work runs through the agent.

Numbers and facts

• −80% — the GPT-5.6 Luna price cut; Sol Fast: up to 2.5× lower latency at 2× the price (OpenAI)
• 40 — Inkling-Small on the Artificial Analysis Intelligence Index, within a point of the flagship (benchmark)
• 30.2% vs 7.8% — Opus 5 vs GPT-5.6 Sol on the official semi-private ARC-AGI-3 with the standard harness; with retained reasoning and compaction Sol reaches 38.3% on the public set (breakdown)
• 400M+ agent memories — Mem0's migration from pgvector to TurboPuffer: 70ms p90 hybrid retrieval, 97% recall@10 (TurboPuffer)
• 2× — end-to-end MI355X performance gain by team Readonflow at GPU Mode's AMD kernel hackathon (SemiAnalysis)

Points of view: agent memory — a product, or still an experiment?

• For: memory is being productized — Perplexity launched Projects with persistent "Brain" memory, and Arav Srinivas positions it as a multiplayer, agentic operating system for work.
• Neutral: long-horizon evals increasingly measure the whole system, not the weights — François Chollet clarified ARC's rules (bespoke harnesses are banned; general-purpose API features are fine if reported), and researchers converge on the view that reasoning retention, truncation policy and tool orchestration matter as much as the model.
• Against: a fresh paper suggests filesystem-style memory stores can halve retrieval cost at scale, yet final answer quality did not improve in the study — and store quality degraded under most management agents.

Tools and techniques

• Don't quantize the KV cache on long context — a practical tip from the local-LLM community: disabling Q8 KV-cache quantization noticeably improved 100k+-context coding with Qwen3.6-27B, while 4-bit weight quants remained fine.
• Prebuilt wheels for GPU-heavy packages — Astral open-sourced its build pipelines for FlashAttention and DeepSpeed wheels: less pain installing heavy dependencies.
• mDenseOn and mLateOn — fully open multilingual retrieval models for long-context and code retrieval; follow-up metrics show especially strong generalization for the late-interaction variant.

In brief

• METR reached agreement with OpenAI and Redwood Research on an independent review of the model behavior in the Hugging Face incident — scope and tentative conclusions to be published.
• Politico reconstructed the incident timeline: 17,600 autonomous actions across July 9–13 and two model instances involved, including an unreleased internal prototype.
• Simon Willison criticizes OpenAI and Anthropic over search-index opacity: the products lean heavily on search, but who actually powers it is only visible in subprocessor listings.
• Ben's Bites: ChatGPT hits 1 billion users.
• The AI Daily Brief breaks down six questions every enterprise has to answer about AI — from token budgets to business-model change.