The Brief: The Cost War Goes Public
GPT-5.6 and Grok 4.5 reached the public within forty-eight hours, both racing on price and agentic coding — the clearest sign yet that the frontier's contest has moved from capability to cost.
GPT-5.6 reaches general availability across ChatGPT, Codex and the API1
We previewed the GPT-5.6 line on the 7th; on the 9th the whole family went generally available — Sol at $5/$30 per million input/output tokens, Terra at $2.50/$15, Luna at $1/$6, each with a roughly 1M-token context window and 128K max output.2 The pitch is long-horizon agentic work, not a benchmark headline; early hands-on reports are mixed on whether it beats Claude Fable 5 at hard coding.
Why it matters → A limited API preview became a full consumer and developer rollout in under two weeks, and the flagship held its $5-in/$30-out price. The contest OpenAI is entering is no longer who posts the top score but who sustains agentic quality at a price the market will pay.
Grok 4.5 goes public as an "Opus-class" model at $2/$63
xAI released Grok 4.5 on 8 July, its first model built expressly for coding and agentic tasks and its first since the company folded in Cursor, on whose sessions it was trained. Musk calls it "Opus-class, but faster, more token-efficient and lower cost"; priced at $2 per million input and $6 per million output tokens, it undercuts Opus 4.8 and GPT-5.5 by more than 60%, and it is not yet available in the EU.
Why it matters → This is the third near-flagship model in ten days to lead with price rather than a capability claim, after Claude Sonnet 5 and GPT-5.6. The competitive front we have called price-performance is now the whole battlefield, and Cursor's session data becoming a training moat is the quieter structural move.
Chinese open models now take up to 46% of US enterprise tokens4
Chinese models have accounted for more than 30% of tokens used by US companies every week since 8 February, peaking near 46%, against an 11% average over the prior year, on Vercel's usage data reported by CNBC on 7 July. The economics are stark: DeepSeek V4 Flash lists at $0.14 per million input tokens versus $5.00 for GPT-5.5, and Z.ai's GLM 5.2 landed within a point of Opus 4.8 on one agentic benchmark at roughly a fifth of the cost. Startup Lindy moved 100% of its traffic from Claude to DeepSeek.
Why it matters → This is the demand side of the same cost war the US launches are fighting: when a task does not need the best model, buyers route it to the cheapest one that is good enough, and open weights out of China are winning that trade. The frontier labs are cutting prices into a market already defecting on price.
A single benchmark run cannot characterise an agent's model discovery, a new paper argues5
In work posted 7 July, He and colleagues treat coding agents as stochastic operators and apply formal experimental design to autonomous model-discovery runs by Codex and Claude Code, varying reasoning effort, task type and training-data composition. Their finding is methodological: an agent's discovery behaviour "cannot be adequately characterised by a single benchmark run," so point-estimate leaderboards can silently misrank systems.
Why it matters → It extends the eval-reliability thread we have tracked — the inference-compute caveat on the 7th, the "Every Eval Ever" schema on the 9th. The variance is not only in the harness; it is in the agent itself, which means an agentic score is a sample, not a measurement.
Brussels defers the AI Act's high-risk obligations by more than a year6
The Council gave final approval on 29 June to the "Digital Omnibus" simplification of the AI Act, with the measure entering into force three days after its imminent publication in the Official Journal. High-risk obligations for stand-alone Annex III systems, due from 2 August 2026, now move to 2 December 2027; those for high-risk AI embedded in regulated products slip from 2 August 2027 to 2 August 2028. A new prohibition on AI-generated non-consensual intimate imagery is added.
Why it matters → We noted on the 9th that the Act's supervisory powers over general-purpose models still arrive on 2 August. That track holds; the high-risk track has slipped to 2027 and 2028. The Act's timeline has forked, and compliance teams now face two very different clocks rather than one deadline.
Meta explores renting out its spare compute, inverting the scarcity story7
Reporting at the start of the month describes "Meta Compute," an internal plan to lease idle data-centre capacity — either raw compute or access to hosted models — to outside clients, following the same move xAI has made. The signal alone moved markets: Meta's shares rose while semiconductor and neocloud names fell on the read that a hyperscaler now has slack to sell.8
Why it matters → It is the supply-side face of this week's price war. If the largest buyers of accelerators are sitting on enough spare capacity to become sellers, the assumption that compute is the binding scarce input — the premise under every capex forecast — is the one to watch break first. This is reporting on an internal plan, not a launch; treat the market reaction as the fact.
Sources
- OpenAI — "Previewing GPT-5.6 Sol: a next-generation model" (model tiers, per-token pricing, agentic framing)
- Simon Willison — "The new GPT-5.6 family: Luna, Terra, Sol" (9 July 2026; general availability across ChatGPT/Codex/API, ~1M context, 128K output, Feb-2026 cutoff, hands-on notes)
- xAI (SpaceXAI) — "Introducing Grok 4.5" (8 July 2026; coding/agentic focus, $2/$6 pricing, Opus-class claim, Cursor training, EU availability)
- CNBC — "Chinese AI models are gaining ground with U.S. companies as OpenAI, Anthropic costs surge" (7 July 2026; >30% and up to 46% of US enterprise tokens, DeepSeek V4 Flash $0.14, GLM 5.2, Lindy switch; Vercel data)
- He, Liu, Kuhlman & Deng — "An Experimental Design Approach to Evaluating Agentic AI's Autonomous Model Discovery," arXiv:2607.06413 (7 July 2026; Codex and Claude Code as stochastic operators)
- Council of the EU — "Artificial Intelligence: Council gives final green light to simplify and streamline rules" (29 June 2026; Digital Omnibus, high-risk dates moved to 2 Dec 2027 / 2 Aug 2028, new NCII prohibition)
- TechCrunch — "Meta, like SpaceX, looks to turn excess AI compute into cash" (1 July 2026; "Meta Compute" plan to lease idle capacity)
- CNBC — "Meta stock pops on cloud push to sell excess AI compute power capacity" (1 July 2026; Meta shares rose while Micron, AMD and Asian chipmakers fell on the Meta Compute report)