AI News
Short, sourced briefs on the AI models and companies we track: what happened, and why it matters.
- Anthropic completes Claude 5.5 — Opus 5.5, Sonnet 5.5 and Haiku 5.5 in 15 days Anthropic shipped Claude Opus 5.5 (Sept 22), Sonnet 5.5 (Sept 28) and Haiku 5.5 (Oct 7): Opus drops to $4 / $20 and 40% lower cost per task than Opus 5, Sonnet keeps $2 / $10, and Haiku falls to $0.10 / $0.50 — about 75% cheaper than Haiku 4.5. Anthropic
- OpenAI rolls out GPT-6 in four steps — Astra, Sol, Luna and a $2 / $10 GPT-6.1 Sol From GPT-6 Astra (Sept 3, $10 / $50) to Sol and Luna (Sept 22, 50% price cuts), GPT-6.1 Sol (Sept 29, near-Astra at one-fifth the price) and a ChatGPT rollout with Intelligent UI on Oct 7, OpenAI's GPT-6 family now spans $0.10 to $50 per million tokens. OpenAI
- Mistral Large 4 enters public preview — 1.05T open-weight MoE, weights due end of October Mistral launched Large 4 in public preview on October 6: 1.05T parameters (52B active), 1M context, native image input, DeepSWE v1.1 61.7%, top-five on the Artificial Analysis Cyber Index — at $1.36 / $4.18 per million tokens, half price during the preview. Mistral
- Google announces Gemini 4 Argon — $2 / $10, a 1M-token output limit and a limited rollout On Sept 30, 2026 Google DeepMind unveiled Gemini 4 Argon, leading its own table on DeepSWE v1.1 (77.9%) and long-video understanding at an introductory $2 / $10 per 1M tokens. Access starts with trusted cyber defenders through the Fairwind Program; broad availability is still pending. Google
- China's open-weight wave: GLM-5.3, Qwen3.8, DeepSeek V4.1-Flash and MiMo-V2.6 in six weeks Between August 12 and September 22, Chinese labs shipped open-weight Qwen3.8-Max, GLM-5.3 and GLM-5.3-Flash, Qwen3.8-Flash-Next, DeepSeek V4.1-Flash and MiMo-V2.6-Pro — mostly MIT or Apache, increasingly multimodal, with inputs from $0.15 per million tokens. Alibaba also said Qwen 4 is in training. Open Weights
- xAI ships Grok 4.7 — a bigger base model at Grok 4.6's $2 / $6 price xAI released Grok 4.7 on September 21: a new, larger base model with a longer RL run, 500K context, and Terminal-Bench 4.0 up from 20.3% to 37.6% — at the same $2 / $6 per million tokens as Grok 4.6, with a 2x-speed Fast variant inside Cursor and Grok Build. xAI
- DeepSeek V4.1-Flash: a 552B open-weight multimodal MoE that beats V4-Pro, from $0.15 / $0.60 DeepSeek released V4.1-Flash on September 10: MIT-licensed weights, 8B-active input and 16B-active output, a KV cache a quarter the size, 1M context, DeepSWE v1.1 74.2% and Terminal-Bench 2.1 90.6% — with API prices of $0.30 / $1.20 at peak and half that off-peak. DeepSeek
- AI week in review: Meta open-sources Muse Glimmer, Grok Imagine Image 2.0 hits #2, Wan 3.0 does document-to-video, and more Meta pivots to Apache 2.0 with Muse Glimmer and Spark 1.2, xAI's image model jumps to Arena #2, Alibaba's Wan 3.0 turns documents into 30-second videos, OpenAI ships GPT-5.6-Cyber, and MAGI-2 open-sources a 114B video MoE. weekly
- AI week in review: Qwen 3.8-Max (2.4T), MiniMax H3 open-weight video, FLUX 3 Video GA, and more A packed week: Alibaba ships Qwen 3.8-Max at $2/$6 with open weights promised, MiniMax open-sources the first major video model with native audio, FLUX 3 Video goes GA with pricing, xAI ships 1080p video, and Google launches robotics AI. weekly
- Anthropic launches Claude Opus 5 — new #1 on AA Intelligence Index at unchanged $5/$25 On July 24, 2026 Anthropic released Claude Opus 5 as its frontier flagship: AA Intelligence Index #1 (61), Frontier-Bench SOTA (43.3%), ARC-AGI-3 30.2% (4× next-best), thinking on by default — all at $5/$25, same as Opus 4.8. Anthropic
- Black Forest Labs launches FLUX 3 — unified image + video model with real-time editing On July 22, 2026 Black Forest Labs released FLUX 3, a unified architecture that generates 4K stills and 10-second 4K video clips from a single model, with real-time interactive editing. Beta API at launch. Black Forest Labs
- Google ships Gemini 3.6 Flash — 17% fewer tokens, lower price than 3.5 Flash On July 21, 2026 Google released Gemini 3.6 Flash as the new workhorse model: 17% fewer output tokens than 3.5 Flash at $1.50/$7.50 per 1M tokens with 1M context. Replaces 3.5 Flash; Gemini 3.5 Pro remains absent. Google
- GPT-5.6 reaches general availability — Sol, Terra, Luna now open to all developers On July 9, 2026 OpenAI made GPT-5.6 generally available after a two-week government-coordinated limited preview. The three-tier family — Sol ($5/$30), Terra ($2.50/$15), Luna ($1/$6) — brings 1.05M context and Sol ultra's Terminal-Bench record (91.91%) to all API customers. OpenAI
- Thinking Machines Lab launches Inkling — 975B open-weight MoE under Apache 2.0 On July 15, 2026 Mira Murati's Thinking Machines Lab released Inkling, a 975B-parameter MoE (41B active) under Apache 2.0. Positioned as a fine-tuning base rather than a frontier competitor, with multimodal input and full-precision weights on Hugging Face. Thinking Machines Lab
- Kimi K3 goes fully open — 2.8T MoE weights, technical report and infrastructure code released On July 27, 2026 Moonshot AI released the full Kimi K3: 2.8T-parameter MoE (104B active), 1M context, open weights on Hugging Face, a 47-page technical report, and infrastructure code on GitHub. API at $3/$15. Custom license (not MIT/Apache). Moonshot
- EU orders Google to open Android AI assistants and Search data under the DMA Two July 16 DMA decisions force Google to give rival AI assistants comparable Android access by July 2027 and share Search data with eligible rivals from January 2027 — a direct hit on Gemini’s default advantage. EU
- Kimi K3 soft-launches with up to 1M context — official specs still pending On July 16, 2026 Kimi's K3 flagship appeared in product surfaces (alongside K2.7 Code) with tiered context up to 1M. Moonshot has not yet posted a full model card, pricing, or benchmarks. Kimi
- xAI launches Grok 4.5 — coding flagship at $2/$6 with 500K context On July 8, 2026 xAI shipped Grok 4.5 as its coding-and-agent flagship: default in Grok Build and Cursor, $2/$6 per 1M tokens, 500K context — trained alongside Cursor. xAI
- xAI open-sources Grok Build — the Rust harness behind its coding agent SpaceXAI released Grok Build under Apache 2.0 on GitHub: agent loop, TUI, tools, and MCP/plugins — with a local-first path via config.toml after a privacy backlash. xAI
- Fable 5 is back: inside the 18 days the US government kept Anthropic's best model offline The US lifted export controls on Claude Fable 5 and Mythos 5 on June 30 after an 18-day suspension triggered by an Amazon-reported jailbreak. Fable 5 returned globally July 1 with a hardened safety classifier — and AI releases gained a new regulatory playbook. Anthropic
- DeepSeek launches its biggest hiring spree yet, aiming to double every department After a reported ~$7.4B raise, DeepSeek opened 33 roles across seven categories to at least double every team — a bet on talent in the race to AGI. DeepSeek
- Mistral releases OCR 4: document AI that understands structure — and runs on your own servers Mistral OCR 4 returns bounding boxes, block types and confidence scores instead of flat text, speaks 170 languages, and ships in a single self-hostable container at $4 per 1,000 pages. Mistral