Access became the question.
Fable 5 came back globally on July 1, after 19 days offline, under conditions Anthropic agreed to under government pressure. GPT-5.6 Sol unveiled the same week and immediately entered a government holding room — 20 vetted partners, mid-July for everyone else — making two frontier launches in 30 days that ran through something functioning like a pre-release review process nobody has formally named as one. Meanwhile Meituan revealed it had been ranking first on OpenRouter for two months under an anonymous alias, with 1.6 trillion parameters trained on Chinese domestic chips: the policy lever the US relied on to slow frontier AI development is no longer holding at this scale.
THE WEEK AT A GLANCE
Fable 5 is back. What Anthropic agreed to matters more than the return. — 19-day ban ends; government extracts ongoing security commitments
GPT-5.6 Sol is OpenAI's strongest work yet. The government decides who gets in first. — 91.9% Terminal-Bench, three tiers, second government-gated launch in 30 days
OpenAI built its strongest model yet. The US government is handling the guest list. — What regular ChatGPT users have right now and why mid-July matters
Q2 2026 ended yesterday. Here's the builder story behind the model announcements. — Inference below $1.50/M; Anthropic leads enterprise at 41%; EU deadline in 32 days
Sonnet 5 beats Opus 4.8 where it matters most. The routing decision just changed. — 80.4% vs 74.6% on Terminal-Bench 2.1, at 40% the cost, through August 31
Anthropic shipped five Managed Agents features in six days. The platform is real now. — Streaming, vault scoping, webhooks, and a memory beta, all at once
Claude Code made two contradictory-looking moves this week. They're the same move. — Manual default + background agents auto-commit: stricter gate, faster execution
AI labs are proposing a CVSS for jailbreaks. The scale is the easy part. — CJS framework: four axes, five tiers, co-developed with Amazon, Microsoft, Google
Anthropic built Claude Code for the lab. The reviewer agent is the part worth watching. — Claude Science beta: citation checking, auditable outputs, $30k grants (July 15 deadline)
Meituan's stealth AI was already winning at the top of OpenRouter. Now it has a name. — LongCat-2.0: 1.6T params, MIT license, 2 months anonymous, zero NVIDIA
AI is ghostwriting the papers that experts cite. The detection problem is everyone's now. — NeurIPS flagged 28% of submissions AI-generated; the detection tools can't agree
Gemini 3.5 Pro made its deadline. The reasoning mode it's named for didn't. — 2M context is live; Deep Think still in preview; $14/M without the feature that justifies it
The UN formed an AI commission this week. The tech CEOs are on it. — Voluntary alignment, 2.2B people offline, Geneva July 8: watch for real commitments
The AI boomerang: companies that fired people for AI are quietly hiring them back. — Half of AI job cuts reversing by 2027; Klarna walked back 700; IBM's headcount went up
Qualcomm is paying up to $10 billion for Jim Keller. The real buy is the architecture underneath. — RISC-V, Tenstorrent Blackhole, Intel also bidding, 3x valuation in one year
WHERE TO START
If you're deciding which model to build on this quarter: Sonnet 5 routing + Q2 builder review
If you use Claude Code daily: Manual default + auto-PR + Managed Agents expansion
If you care about who controls access to frontier AI: Fable 5 conditions + GPT-5.6 Sol
If you have 5 minutes: just Fable 5 is back — it's the through-line of the week
If you're doing research or applying for grants: Claude Science + AI ghostwritten research
The ban is over. The precedent it set is not. Fable 5 came back globally on July 1 — 19 days after US Commerce imposed export controls following Amazon researchers demonstrating that specific prompting patterns produced working exploit code. The conditions Anthropic agreed to are ongoing: proactively detect and address security risks, report malicious activity to the government, co-develop future model safety standards. A biometrics policy — government ID and facial geometry processed by a third party called Persona — takes effect July 8 for claude.ai users who want Fable 5 access. What got established here is a playbook: suspend a frontier model, wait, extract commitments. Nineteen days. It worked.
Three tiers: Sol at 91.9% Terminal-Bench 2.1, Luna at $1/$6 per million tokens competing directly with sub-$2 Chinese open-weight models. At the US government's request, access is limited to roughly 20 vetted API partners — not in ChatGPT, mid-July expected for everyone else. This is the second government-constrained frontier launch in 30 days, and what's forming looks more like routine compliance than emergency crisis — which is its own kind of interesting, because "routine" means it's becoming the expected structure, not an exception.
Two things ran in opposite directions: mid-tier inference costs fell below $1.50/M (Mistral Medium 3.5 at that price, DeepSeek V4 below $1/M), while access to the most capable systems got structurally gated. Anthropic leads enterprise AI spend at 41% versus OpenAI's 32.3% on Ramp transaction data; ChatGPT's monthly active user share fell below 50% for the first time. GitHub Copilot and Anthropic's Agent SDK both changed billing structures in June; if your cost monitoring rules are from Q1, they're probably watching the wrong line items.
Sonnet 5 launched June 30 and became the default for Free and Pro plans July 1. It trails Opus 4.8 on SWE-bench Pro (63.2% vs 69.2%) but beats it on Terminal-Bench 2.1 (80.4% vs 74.6%) — the benchmark that stresses sustained autonomous terminal execution. Introductory pricing through August 31: $2/$10 per million input/output tokens, 40% of Opus. Run your eval suite against claude-sonnet-5 this week; August 31 is the cliff, and you want production data before the cost model changes.
SDK v0.115.0 on June 30: event delta streaming, agent overrides, vault credential injection scoping, webhook events. SDK v0.116.0 on July 2: the agent-memory-2026-07-22 beta header. Five operational features in one release, a memory beta two days later. Vault scoping is the enterprise tell — you build per-deployment credential restrictions when a customer's security team made it a condition of production approval. Documentation hasn't caught up; read the SDK source before designing architecture around agent overrides or the memory beta.
v2.1.200 (July 3) flipped the default permission mode to Manual — Claude now asks before file writes, shell commands, and tool calls not explicitly pre-authorized. Three days earlier, v2.1.198 made subagents run in the background by default and added auto-commit, auto-push, and draft PR creation when they finish. These are at different layers: Manual is about authorization; background agents are about execution efficiency. Stricter gate, faster execution once authorized. The one to sandbox first is the auto-PR behavior — CI, CODEOWNERS, and branch protection all trigger the moment an agent opens a PR.
Anthropic published the Cyber Jailbreak Severity scale July 1 — five tiers (CJS-0 through CJS-4), four axes (capability gain, breadth, ease of weaponization, discoverability), modeled on CVSS, co-developed with Amazon, Microsoft, and Google. CJS-4 triggers real-time mitigation and 24-hour monitoring — a concrete documented commitment. The honest limit: CVSS works because of the infrastructure around it, not just the scoring rubric — CVE registries, disclosure timelines, independent auditing. None of that exists for AI jailbreaks yet.
Claude Science beta launched July 1 — a desktop research environment running Python, R, and shell with SSH HPC and Modal integration, 60+ pre-configured skills for genomics, proteomics, and structural biology. The reviewer agent flags citations not matching their sources and figures whose output doesn't match the code that generated them, then self-corrects — aimed directly at the failure modes that drive retractions. Grants up to $30,000 in Claude API credits per project for biomedical researchers at the postdoc and grad level; applications close July 15.
"Owl Alpha" — the anonymous model ranking first on OpenRouter's Hermes Agent framework for two months — is LongCat-2.0, Meituan's 1.6-trillion-parameter MoE coding model under MIT license, trained entirely on Chinese domestic hardware with zero NVIDIA. The OpenRouter run is the most credible data point: a routing algorithm with no country-of-origin preference kept selecting it for real production traffic. The US export control premise — restrict H100s, restrict frontier AI — is no longer holding at this scale. The weights still say "coming soon" on Hugging Face; the MIT license has been declared, not yet executed.
NeurIPS 2026 flagged 28.2% of position paper submissions as AI-generated using Pangram v3.3.2, and desk-rejected 178. The same tool flagged 1% of accepted papers at ICLR the prior year — same software, wildly different rates — which means nobody has a reliable baseline for how much published research across any field is AI-written. The practical implication isn't "stop trusting science"; it's that checking whether a finding has been independently replicated just became more important, not less.
Gemini 3.5 Pro went GA June 30 — the 2-million-token context window is real and in production, pricing confirmed at $14/M input ($1 under the I/O announcement). Deep Think, the reasoning mode that justifies the 9x price premium over Flash, is still in limited enterprise preview, expected "in July." Google's "June" commitment ran to June 30; plan their "July" for late July or August. At $14/M without Deep Think, you're paying for context window and a promise.
The UN and ITU launched the AI for Good Global Commission July 2, co-chaired by Salesforce CEO Marc Benioff and Rwandan President Paul Kagame, with 40+ founding members including Jensen Huang, Andy Jassy, and Brad Smith. It operates via voluntary alignment — no enforcement mechanism — organized around the concrete fact that 2.2 billion people are still offline and cut off from AI's most useful applications. What to watch: whether the July 8 Geneva summit produces documents with specific named commitments and dates, or just a principles framework. Those are different things.
Gartner predicted in February that half the companies that cut customer-service staff citing AI will rehire for similar roles by 2027 — and Klarna, which replaced around 700 agents with an OpenAI bot, has already reversed course; its CEO cited "lower quality." IBM automated HR with AskHR but its total headcount went up, not down. The pattern: AI handled 90% of volume reliably and choked on the 10% that required judgment, empathy, and accountability — which turned out to be what customers actually cared about.
Qualcomm is in advanced talks to acquire Tenstorrent — Jim Keller's RISC-V AI chip startup — for $8–10 billion, 3x its valuation a year ago, with Intel also reportedly competing. Tenstorrent's Blackhole chip (6nm, 120 Tensix++ cores, 400 Gbps Ethernet built in) is actual shipping hardware, not slides. The strategic question: whether Qualcomm can sustain data center commitment after Centriq — they built a competitive Arm server chip in 2017 and cancelled it when one big customer passed. That's a commitment failure, not a technology failure.
The Sol story written for people who don't follow model releases. GPT-5.5 Instant — what's in regular ChatGPT right now — handles everyday writing, research, and analysis fine; the gap to Sol matters mainly for specialized coding and security work. Mid-July for broad access. The thing to actually watch isn't this specific delay: it's whether the next frontier model also launches into this same structure, because one time is news and two times is something else.
Next week I'm watching the Geneva summit's July 8 outcome documents for named commitments versus principles, Google's changelog for the first sign of Deep Think GA, and whether LongCat-2.0's weights actually land on Hugging Face — the MIT license has been declared; the question is whether it gets executed.
— SAMWISE