AI News — July 25, 2026

Model releases

• [2026-07-24] Anthropic — Claude Opus 5 ships as the new default Opus: roughly Fable 5-level intelligence at half Fable’s price, with a per-request effort dial. Anthropic released Claude Opus 5, live across Claude.ai, the API, Claude Code, and Cowork, now the default on Claude Max and the strongest model on Pro. Standard pricing is $5 / $25 per million input/output tokens — identical to Opus 4.8 and half of Fable 5’s $10/$50 — with a fast mode running ~2.5× faster at double the base rate ($10/$50); the context window is 1M tokens. The headline feature is a per-request effort control (low/medium/high, extending to xhigh/max) that trades token spend for reasoning depth, and Anthropic says Opus 5 converts extra effort into better results more reliably than any earlier Opus. On its own numbers it lands around Fable 5-level intelligence at half the price, sets a new state-of-the-art on agentic coding and knowledge-work benchmarks (Anthropic cites 96.0% SWE-bench Verified, 79.2% SWE-bench Pro, and a jump to 59.4% on SWE-bench Multimodal), and more than doubles Opus 4.8 on agentic terminal coding, though it trails the Mythos line on cybersecurity. Anthropic also frames it as its safest Opus: ~85% lower safety-classifier trigger sensitivity than Fable 5, the lowest misalignment score it has measured, and no data-retention requirement for general access (unlike Mythos-class models). It matters because it inverts the usual flagship pattern — a stronger frontier model arriving at the same Opus price and below Fable’s, with a cost/capability dial — resetting what the top tier costs to run rather than raising it. (official, TechCrunch, Fortune, Bloomberg)

  Opus 5 landing at Opus 4.8’s exact $5/$25 base makes the upgrade a swap rather than a budget approval — the model under Claude Max changed today with no new rate to sign off on, so the one genuinely new decision it introduces is whether and when to reach for the effort dial.

  For Software Developers: With claude-opus-5 now the default in Claude Code, the low→medium→high dial (extending to xhigh/max) turns spend into a per-task choice on a single model: keep routine edits and boilerplate at low, and save high or above for a thorny refactor or a deep debugging session — no more switching to a cheaper model just to control cost, since the base rate is the same either way.

Coding agents

• [2026-07-24] Anthropic — Claude Code 2.1.219–2.1.220 make Opus 5 the default and reverse course on nested subagents. 2.1.219 adds Claude Opus 5 (claude-opus-5) as the new default Opus model (1M context; fast-mode pricing $10/$50), points /fast at Opus 5 and Opus 4.8, and drops Opus 4.7 from fast mode. Notably it raises the default nested-subagent spawn depth from 1 back to 3 (CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH) — a partial reversal of 2.1.217’s no-nested-subagents default from three days ago (last briefing) — so subagents can again spawn their own subagents out of the box. It also adds a sandbox.network.strictAllowlist setting (deny non-allowlisted hosts for sandboxed commands without prompting), a DirectoryAdded hook that fires when /add-dir registers a new working directory mid-session, mcp_server_errors in the headless init event, and --forward-subagent-text for nested-subagent output in stream-json. 2.1.220 is a bug-fix/reliability release. It matters because the fan-out default just moved the opposite direction from last week’s hard limits — teams that pinned a version expecting the depth-1 default should note the change before updating. (official, prior coverage)

  Within one week the nested-subagent default has flipped twice — 2.1.217 killed it three days ago, 2.1.219 restores depth 3 — so for anything automated the real lesson isn’t which default is right but that fan-out behavior now changes release to release, which makes a pinned version the only stable contract.

  For Security Engineers: sandbox.network.strictAllowlist denies any non-allowlisted host for sandboxed commands without prompting, so an unattended or CI agent can now be locked to a fixed egress set as a hard default — closing the interactive-approval gap where a sandboxed command could still reach an unlisted endpoint on a one-off OK.

• [2026-07-22] Cursor — Cursor Router lands as “Auto” mode: per-request model routing with a cost/quality dial. Cursor shipped Auto mode powered by Cursor Router, which classifies each request by task type and complexity and routes it to the model it judges best, with three optimization modes — Cost, Balance, and Intelligence — plus admin controls for per-team defaults, mode restrictions, and model allow/block lists. Cursor says A/B tests across millions of requests delivered “frontier-quality performance at 60% savings.” It’s available across desktop, web, iOS, CLI, and the SDK, and Cursor also opened Cursor for iOS in public beta on all paid plans. It matters because request-level routing is becoming the default cost lever for coding agents — the same “trade spend for capability per call” idea Opus 5’s effort dial exposes, here handled by the tool instead of the model. (official, blog)

  Cursor putting routing behind Cost/Balance/Intelligence modes with per-team admin defaults turns model choice into a policy an org sets once rather than a per-developer habit — though the “60% savings” is Cursor’s own A/B figure, not an independent benchmark, so it’s worth validating on your own request mix before banking on it.

  For Solution Architects: Picture standardizing an org on Cursor — set Balance as the default mode, restrict which optimization modes teams can pick, and maintain a model allow/block list, and the cost/quality posture holds across every surface (desktop, web, iOS, CLI, SDK) instead of each developer choosing a model per request.

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)

  Two days out, the move that pays off is lining up an A/B now: keep a handful of real tasks ready to replay against a third-party K3 endpoint the moment one appears, since the first independent evals — not Moonshot’s — are what will actually place it against a freshly repriced Opus 5 and Gemini 3.6 Flash.

• Gemini 3.5 Pro — still undated even as the Flash line ships. With 3.6 Flash and its siblings out (July 21) while 3.5 Pro remains without a date or model card (“coming soon” on the Gemini model page), the “Flash ships first” tell the last several briefings tracked 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)

  Opus 5 shipping today at a lower flagship price sharpens what the delay costs Google: the tier 3.5 Pro is meant to contest keeps moving while it stays dateless, so the item to watch is unchanged — a Google post with a real date and model card — but the price of the silence is now concrete.

• MCP spec finalization (July 28) — the 2026-07-28 release candidate (stateless core, Extensions framework, Tasks, MCP Apps, authorization hardening) is in its final 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)

  Three days out, the RC is close enough that a late Mcp-Method/Mcp-Name header change would itself be the news — so the practical posture now is to freeze integration testing against the current beta headers and watch the changelog for a diff, rather than expecting more structural change before the lock.