AI News — July 23, 2026

Model releases

• [2026-07-21] Google — Gemini 3.6 Flash lands with two siblings — 3.5 Flash-Lite and a gated 3.5 Flash Cyber — but still no 3.5 Pro. 3.6 Flash is the new “workhorse” tier: Google says it uses ~17% fewer output tokens than 3.5 Flash on the Artificial Analysis Index — up to 65% fewer on DeepSWE (by Datacurve) — at a lower cost per output token, while scoring higher on every internal eval. Artificial Analysis clocks the high-reasoning variant at 50 on its Intelligence Index (tier median 31) and ~275 tokens/sec, with pricing around $1.50 / $7.50 per million input/output tokens ($0.15/M cached) and the knowledge cutoff jumping from Jan 2025 to March 2026. It’s live for developers today in the Gemini API via Google AI Studio and Android Studio. 3.5 Flash-Lite is the fastest, cheapest 3.5-class model (~350 tokens/sec) and is rolling into Google Search; 3.5 Flash Cyber, tuned to find and fix vulnerabilities, is deliberately gated to governments and trusted partners via CodeMender under a limited-access pilot, citing the tech’s dual-use risk. It matters because after three missed 3.5 Pro targets and a “months behind schedule” report, this is the first concrete evidence Google’s Gemini pipeline is shipping again — the “upgraded Flash could land first” scenario the last several briefings tracked — and a cheaper-yet-stronger Flash resets the cost/quality floor for the workhorse tier. (official, benchmarks, source)

  Because the ~17% token cut and the lower per-token price stack rather than trade off, the effective spend drop for a high-volume workload is bigger than either figure alone — and it compounds hardest exactly where output dominates, like long agentic loops or batch generation, which is the workload a Flash tier is meant to carry in the first place.

  For Software Developers: If you drive a coding agent through the Gemini API, the DeepSWE number is the one to actually measure — at the 65% end of that “up to” range, a SWE-style task that burned a dollar of generated tokens on 3.5 Flash lands near thirty-five cents on 3.6 Flash before the lower per-token rate is even applied. Point a throwaway branch’s agent runs at gemini-3.6-flash, replay a couple of real tickets, and compare the token bill against your current default before flipping it in production.

Coding agents

• [2026-07-22] Anthropic — Claude Code 2.1.218 moves /code-review to a background subagent and fixes a Windows path-corruption bug. After 2.1.215 made /code-review a manual step and 2.1.216–2.1.217 added subagent fan-out limits (last briefings), 2.1.218 runs /code-review as a background subagent so review work no longer fills the main conversation (and it keeps stacked slash commands as its review target). It also fixes Windows paths with \u-prefixed segments (e.g. C:\Users\unicorn) being mangled into CJK characters — which had made those files inaccessible to tools — adds a confirm step so the left-arrow key can’t silently discard a conversation with no undo, and surfaces HTTP status and error text in claude mcp list and /mcp when a server fails to connect. It’s a routine point release, but the code-review-as-subagent change is worth noting for anyone who scripts a review pass and was relying on it running inline. (official)

  The Windows \u fix is narrower than it sounds but brutal if it caught you: any repo checked out under a path segment starting \u — \Users, \updates, \unicorn — had those files silently turned into CJK gibberish and made unreachable to tools, so a phantom “file not found” on a path that plainly exists is the symptom this release finally explains and closes.

  For Software Developers: If you’d wired /code-review into a script that runs it and then greps the main transcript for its findings, 2.1.218 quietly breaks that read — the review now runs in a background subagent, so its output lives in that subagent’s context rather than the main conversation. The adjustment is to have the review emit findings to a file or a structured output the parent step reads, instead of scraping them out of the main log the way an inline pass let you.

Watch list

• Kimi K3 open weights (July 27) — Moonshot has K3 live via API (kimi-k3, $3/$15 per million) but says the 2.8T weights go public July 27, reportedly under a modified-MIT-style license (unconfirmed until the model card ships). Watch for the model card, the license’s rumored monthly-active-user clause, whether third-party inference providers stand up endpoints undercutting Moonshot’s own pricing, and the first independent, non-arena evals. (prior coverage)

  Worth watching less for the weights than for the license fine print: a monthly-active-user clause bolted onto a nominally MIT-style license would quietly disqualify K3 as a base model for any product past that user count, so if you’re evaluating it to build on, the license text is the thing to read first — before the eval scores that will dominate the coverage.

• Gemini 3.5 Pro — still undated even as the Flash line ships. With 3.6 Flash and its siblings now out (July 21) while 3.5 Pro remains without a date or model card, the “Flash ships first” tell the last several briefings watched has played out — the pipeline is clearly moving, but the flagship is still missing. The headline specs (2M-token context, Deep Think, pricing) remain reported, never confirmed. The signal worth revising a plan around is a Google post carrying an actual 3.5 Pro date and a published model card, not another leaked target. (unconfirmed) (report, prior coverage)

  Now that Flash has actually shipped, the missing-Pro question narrows in a useful way: the pipeline clearly isn’t frozen, so a still-dateless flagship points at something specific to the Pro tier — capability targets it can’t yet hit, safety review, or compute allocation — rather than a company-wide stall, which is a sharper thing to reason about than another rumored window.

• MCP spec finalization (July 28) — the 2026-07-28 release candidate (stateless core, Extensions framework, Tasks, MCP Apps, authorization hardening) is in its validation window, with the Python/TypeScript/Go/C# SDKs updating against it in beta. Security analysts continue to flag the stateless rework’s new surface — client-held state manipulation and a “hit-and-run” async-task DoS — on top of the Mcp-Method/Mcp-Name routing headers. Watch which of the four SDKs reaches parity last and any final header changes before the RC locks. (official, prior coverage)

  For anyone already running an MCP server, this five-day validation window is the cheap time to catch breakage: the new Mcp-Method/Mcp-Name routing headers and the stateless core almost certainly need changes to server-side request handling, and finding that against a beta SDK now beats discovering it the day the RC locks and clients start speaking the final protocol.