← All news

Weekly recap

AI News — Week of July 13–19, 2026

#weekly

Coding became the axis that reshuffled the field: Google's Gemini 3.5 Pro was confirmed months late for coding shortfalls (~$200B off Alphabet) while China's Moonshot took #1 on an external frontend-coding board with the largest open-weight model ever announced.

The week in brief

This was the week coding stopped being one axis among many and became the axis that reshuffled the frontier. Google’s Gemini 3.5 Pro missed its July 17 target and was confirmed months behind schedule specifically on coding — a slip that cost Alphabet roughly $200 billion in market value — while in the same window China’s Moonshot took #1 on an external frontend-coding leaderboard with Kimi K3, the largest open-weight model ever announced. Around that, Anthropic bought Fable 5 one last included week before its metered cutover, and the tooling side spent the week hardening the approval surfaces agents run on.

Biggest stories

  • Gemini 3.5 Pro confirmed delayed — Bloomberg reports it’s “months behind” on coding, and Alphabet sheds ~$200B. A day before its widely reported July 17 GA date, Bloomberg reported Google is months behind on its most powerful model because coding capabilities fell short of internal expectations; Google confirmed no GA, and Alphabet closed down ~4.4%. The last unlaunched summer flagship landed not as a launch but as an open-ended slip — on the exact axis Fable 5, GPT-5.6, and Grok 4.5 already compete. (unconfirmed specs) (briefing, report)
  • Moonshot’s Kimi K3 arrives as the largest open-weight model ever announced and takes #1 on the external Frontend Code Arena. A 2.8-trillion-parameter (~50B-active) MoE with a 1M-token context, K3 topped Arena.ai’s human-preference frontend board at 1,679 — ahead of Fable 5 (1,631) and GPT-5.6 Sol (1,618) — winning six of seven domains, though its own benchmark table still trails both flagships elsewhere. It’s a proprietary API today at $3/$15; the weights are slated to go public July 27. (unconfirmed weights date) (briefing, source)
  • Anthropic extends Fable 5’s included window a second time — through July 19 — and slips the meter to July 20. Fable 5 was set to leave Pro/Max/Team/eligible Enterprise plans at 11:59pm PT July 12; instead Anthropic kept it included (with the 50% Claude Code weekly-limit boost) through July 19, then moves it to usage credits at $10/$50 per million tokens. Anthropic frames the meter as temporary pending capacity, with no return date. (briefing, official)
  • Claude Code 2.1.211–2.1.212 close a plan-mode permission bypass and add runaway-loop caps. 2.1.212 fixes plan mode auto-running file-modifying Bash (touch, rm) with no permission prompt or canUseTool callback, plus a worktree-symlink escape that could write outside the repo; it adds session-wide WebSearch and subagent-spawn caps. 2.1.211 neutralizes bidi-override and look-alike characters in permission previews so tool inputs can’t visually alter what you approve. (briefing, changelog)
  • GitHub’s /security-review moves out of the CLI into the Copilot app. The AI vulnerability scan GitHub shipped to Copilot CLI in June now runs inside everyday Copilot coding workflows in public preview across Free, Pro, Business, and Enterprise — putting pre-merge scanning on in-flight diffs, before the PR and before CI, in front of the broader Copilot base. (briefing, official)

By area

  • Model releases — Moonshot’s Kimi K3 (2.8T/~50B-active MoE, 1M context, #1 Frontend Code Arena, weights promised July 27) was the week’s one shipped flagship; Google’s Gemini 3.5 Pro missed July 17 and was confirmed months-late on coding, taking ~$200B off Alphabet.
  • Coding agents — GitHub folded /security-review into the Copilot app, and Claude Code 2.1.211–2.1.212 closed a plan-mode permission bypass and added runaway-loop caps — both fixes landing on the approval/confirmation surface rather than the model.
  • MCP — Quiet as shipped work: the July 28 release candidate (stateless core, Extensions, Tasks, MCP Apps, auth hardening) stayed in its validation window, with the new Mcp-Method/Mcp-Name routing headers still flagged as fresh attack surface.
  • AI cost tracking & telemetry — Pricing was the recurring thread: Fable 5’s included window ran through July 19 before flipping to $10/$50 credits on July 20, while Kimi K3 landed at Sonnet-tier $3/$15 for a model Moonshot frames as Opus-4.8-class.

Themes

  • Coding is now the axis that moves the field. Google fell behind on it and lost ~$200B in a day; Moonshot led on it and put a Chinese open-weight model atop a US-dominated frontend board. The differentiation summer buyers weigh most has shifted from raw capability or token efficiency to head-to-head coding — the exact ground the flagships already fight over.
  • Agent approval surfaces are being treated as attack surface. Claude Code’s plan-mode bypass fix and bidi/homoglyph neutralization, plus GitHub moving pre-merge security review into the app, all target the human-in-the-loop confirmation step — a reminder that in agentic tools the approval UI itself is where a skipped prompt or a spoofed preview does the damage.

Still watching

  • Fable 5 metered cutover (July 20) — the second extension ends this week; after 11:59pm PT July 19, included access gives way to $10/$50 credits, so the open question is whether standard Enterprise seats can enable credits at all before the included allowance empties mid-session. (latest, official)
  • Kimi K3 open weights (July 27) — at 2.8T parameters “open” is gated by hardware, so the near-term signal is less self-hosting than whether hosted providers stand up endpoints undercutting $3/$15, and whether the license carries a monthly-active-user clause. Watch for the model card and first non-arena evals. (unconfirmed) (latest)
  • Gemini 3.5 Pro — delayed, no new date — the target that anchored the watch list for over a week is gone with no published replacement or model card; reporting points to an upgraded Flash model in partner testing that could ship first. (unconfirmed) (latest, report)
  • GPT-5.6’s first independent benchmarks — every Sol/Terra/Luna figure still traces to OpenAI; the missing piece is a neutral eval placing Sol on SWE-Bench Pro and DeepSWE next to Opus 4.8 and Grok 4.5. (latest)
  • MCP spec finalization (July 28) — the RC locks next week; the practical gate is whether the Python/TypeScript/Go/C# SDKs all reach stable support against the final Mcp-Method/Mcp-Name headers before it does. (latest, official)