AI News — Week of July 6–12, 2026

The week in brief

This was the week the frontier race became a price race. Three cheaper near-frontier models landed in seven days — OpenAI’s GPT-5.6 (Sol/Terra/Luna) went broadly available, xAI shipped Grok 4.5 at $2/$6, and Meta put its first-ever paid model, Muse Spark 1.1, behind a meter at $1.25/$4.25 — all bracketing Anthropic’s Fable 5 as it flipped to a record-high $10/$50 metered rate, and all validating the cost pressure CNBC measured as US buyers routing a third of their tokens to cheap Chinese open-weight models.

Biggest stories

GPT-5.6 goes broadly available, ending the government-restricted rollout — the US Commerce Department cleared a wide launch under Washington’s new frontier-model oversight framework, lifting the ~20-partner vetted-preview gate that had held Sol, Terra, and Luna since June 26. A government clearance, not a capability milestone, was the gate — the same pattern as Fable 5’s return. (briefing, news)
xAI’s Grok 4.5 ships at $2/$6, trained in partnership with Cursor on real developer-session data — 4th on the Artificial Analysis Intelligence Index (behind only Fable 5, GPT-5.5, and Opus 4.8) at more than 60% below Opus 4.8 and GPT-5.5, turning the frontier race back toward price. (briefing, official)
Meta starts charging for its own model — Muse Spark 1.1 launched on the new paid Meta Model API at $1.25/$4.25 (a quarter of Opus 4.8 and GPT-5.5), topping the MCP Atlas tool-use benchmark (88.1) but trailing the flagships on coding. It marks Meta’s turn from open-weight Llama toward a metered, agent-first product. (briefing, official)
Fable 5’s metered cutover — after a five-day extension, Anthropic’s Fable 5 stayed included at up to 50% of weekly limits only through 11:59pm PT July 12, then moves to usage credits at $10/$50 per million tokens — double Opus 4.8 and the steepest rate Anthropic has published, setting an unusually high frontier-tier floor. (briefing, official)
Work agents break past developers — OpenAI shipped ChatGPT Work (a GPT-5.6-powered agent that acts across apps and files for hours, governed like Codex with an Auto-review gate) days after Claude Cowork reached web, iOS, and Android with cloud execution. Both ship governance-first. (Work, Cowork)

By area

Model releases — Three near-frontier models in one week: GPT-5.6’s broad GA (Sol $5/$30, Terra $2.50/$15, Luna $1/$6), Grok 4.5 ($2/$6), and Meta’s Muse Spark 1.1 ($1.25/$4.25). OpenAI also advanced voice — gpt-realtime-2.1 (reasoning on the Realtime API, −25% p95 latency) and the full-duplex GPT-Live, which replaces ChatGPT’s Advanced Voice Mode.
Coding agents — ChatGPT Work and Claude Cowork (now web/mobile, cloud-executed) both pushed general work agents past developers; Cursor 3.11 added durable, at-mentionable “side chats”; Claude Code shipped v2.1.200–202 and turned auto mode on by default across Bedrock, Vertex AI, and Foundry (with Opus 4.8 on Bedrock); Kiro extended MCP OAuth to strict enterprise servers.
MCP — Mostly the July 28 release-candidate validation window (stateless core, Tasks, MCP Apps, authorization hardening); Kiro’s OAuth-to-strict-servers (client secrets, custom callback paths, skip-DCR) was the week’s one shipped MCP change.
Agent frameworks & interop — Quiet as a standalone area; the interop story rode inside the launches — Muse Spark 1.1’s native primary-agent/subagent orchestration over MCP.
AI cost tracking & telemetry — The dominant thread all week: Fable 5’s $10/$50 metered cutover, OpenAI workspace agents leaving free preview into credit metering (July 6), and CNBC’s finding that Chinese open-weight models now hold 30–46% of US tokens on OpenRouter as frontier prices climb.

Themes

The frontier race became a price race. Three near-frontier coding/agent models arrived in seven days — GPT-5.6’s GA, Grok 4.5, and Muse Spark 1.1 — all landing well under the top-tier rate, all pitched at coding and agentic work. Against Fable 5’s record $10/$50 floor and CNBC’s measured flight to cheap Chinese open-weights, the summer’s competitive axis shifted decisively from raw capability toward cost per token.
Always-on AI moved onto meters, and general work agents grew guardrails. Workspace agents left free preview and Fable 5 flipped to per-token credits — the same flat-plan-to-metered shift GitHub Copilot made in June. In parallel, ChatGPT Work and Claude Cowork both shipped governance-first (Codex-style network controls, Auto-review), because an agent acting unattended across files and apps is exactly where an unscoped permission becomes an exfiltration path.

Still watching

Gemini 3.5 Pro (GA) — with Grok 4.5, Muse Spark 1.1, and GPT-5.6 all now shipped, it is the only major summer flagship still entirely unlaunched: no model card, benchmarks, or pricing for the promised 2M-token context and Deep Think mode. Reporting points to a July 17 target tied to a full architectural rebuild; Google has confirmed none of it. A 2M-token, Deep-Think GA would reset the context-window and reasoning baseline the rest of the field is now priced against. (unconfirmed) (latest)
GPT-5.6’s first independent benchmarks — every performance figure on Sol/Terra/Luna still traces to OpenAI. Grok 4.5 and Muse Spark 1.1 now put competitor numbers on the same coding suites (SWE-Bench Pro, Terminal-Bench, DeepSWE), so the first neutral run won’t just measure Sol — it will rank it head-to-head against the summer’s other flagships. (latest)
Fable 5 behavioral residuals — with pricing settled, the open questions are whether standard Enterprise seats (which carry no included allowance) can enable usage credits at all, and whether Fable 5’s government-coordinated cyber classifier over-blocks legitimate red-team work in steady-state paid use. (latest)
MCP spec finalization (July 28) — the stateless-core rework is breaking for existing servers, so any last change to the new Mcp-Method/Mcp-Name HTTP headers that security analysts flagged as fresh attack surface lands as migration work for every SDK consumer. Watch the remaining SDK support landing. (latest, official)