SpaceXAI released Grok 4.5, OpenAI moved GPT-5.6 from restricted preview to general availability, and Meta launched Muse Spark 1.1 alongside its first paid hosted model API, all within the same week in July 2026. Every lab positioned its release around agentic reasoning, coding, and autonomous task execution rather than conversational quality.

What Each Lab Shipped

Grok 4.5 arrived priced at $2.00/$6.00 per million tokens (input/output), explicitly marketed for coding and agentic tasks. On the Artificial Analysis Intelligence Index, it scores 54, placing it in the same competitive bracket as GPT-5.5 and Claude Fable 5. According to Towards AI, SpaceXAI and Cursor jointly trained the model. Snorkel measured a 29% full-rubric pass rate across roughly 2,000 professional tasks, ahead of GPT-5.5 and Opus 4.8, though Cursor flagged a contaminated CursorBench result and Artificial Analysis recorded a 54% hallucination rate on AA-Omniscience.

GPT-5.6 reached general availability on July 9 with three tiers: Sol (frontier), Terra (balanced), and Luna (efficient). Sol scores 59 on the Intelligence Index, within one point of Claude Fable 5’s 60. On Agents’ Last Exam, an evaluation of long-running professional workflows, Sol scored 53.6, beating Fable 5 by 13.1 points, according to OpenAI. Demand surged immediately: OpenAI reported temporarily removing the five-hour usage limit on Sol pricing tiers due to load from Codex and ChatGPT Work users.

Meta Muse Spark 1.1 launched with the Meta Model API in public preview, priced at $1.25/$4.25 per million tokens with $20 in free credits. This is Meta’s first paid hosted model API, supporting both OpenAI and Anthropic wire formats. It scores 51 on the Intelligence Index, matching GPT-5.6 Luna at a comparable price point.

The Cost-Per-Task Shift

The compressed launch window exposed a pricing dynamic that matters more than token-level pricing: cost per completed task. According to Towards AI, Grok 4.5 charges more per output token than the leading open-weight model GLM-5.2, but uses roughly 14,000 output tokens per benchmark task compared to GLM’s 43,000. The result: Grok 4.5 costs approximately $0.31 per Intelligence Index task versus GLM-5.2’s $0.37, beating open weights on both score and effective cost.

GPT-5.6 Luna and Muse Spark 1.1 tie GLM at 51 on the Intelligence Index while costing about $0.21 and $0.26 per task respectively. Closed models now match or undercut the open-weight frontier on cost per unit of intelligence, which had been open weights’ clearest selling point.

Agentic Positioning Across the Board

The marketing language across all three releases converges on the same thesis: these models are for agents, not chat. Grok 4.5’s materials reference “coding and agentic tasks.” GPT-5.6’s demand surge came from Codex (an agent harness) and ChatGPT Work (autonomous execution mode), not casual conversation. Meta priced its API to compete directly for agent fleet infrastructure, offering wire-format compatibility that lets developers swap models without rewriting integration code.

OpenAI reported over 8 million active users across Codex and ChatGPT Work following the GPT-5.6 GA rollout, up from 5 million weekly active Codex users in early June, per Towards AI. More than 1 million were using Codex for non-software tasks before the Work launch.

What This Pricing War Means for Agent Operators

For teams running agent fleets, the simultaneous launches create real optionality. Meta’s wire-format compatibility with OpenAI and Anthropic APIs means switching costs between providers approach zero for standard inference. The effective metric is no longer tokens-per-dollar but successful-tasks-per-dollar, and three labs are now competing on that measure simultaneously. Agent operators who locked into single-provider contracts earlier in 2026 now face a market where frontier-grade intelligence is available from at least four vendors (including Anthropic) at converging price points.