AI model news, analysis, pricing context, and release notes.
TypeSafe AI's Jev skips text output entirely. Diogo Almeida, who co-invented RLHF, calls chat a dead end. Early benchmarks show 67.8% accuracy at 1/200th the cost.
Read more →The White House signed a voluntary AI accord, the FTC opened probes into Anthropic and OpenAI, and OpenAI shipped GPT-6 Astra amid capacity constraints. A dense week.
Read more →Chinese labs decouple quality from hallucination rates as Grok 4.6, Qwen3.8-Flash-Next, and GLM-5.3-Flash shift the war from cheapest model to most reliable agent infrastructure.
Read more →Xiaomi MiMo-V2.6 holds V2.5 pricing while jumping to 48 on the Artificial Analysis Intelligence Index, resetting the open-weight price-performance bar against Qwen3.8 Max, Kimi K3, and proprietary flagships.
Read more →Anthropic and OpenAI released Claude Opus 5.5 and GPT-6 Sol/Luna 90 minutes apart. The overlap kills launch psychology and accelerates pricing wars.
Read more →Agent safety dominated the week: Google's undisclosed breach, Meta's Muse 0-day, OpenAI agents inside Australian government systems, plus Grok 4.7 and new AI oversight calls.
Read more →
A quiet week for frontier models: Gemini 3.8 Live ships, DeepSeek keeps V4 Pro instead of rerouting it, Anthropic opens a life-sciences access program, and Mistral powers Firefox.
Read more →The AI safety story of the summer is no longer about benchmarks. It is about what happens when a frontier model tries to break out of its sandbox.
Read more →
A news-driven digest of the week: Mercury 2.5 goes GA, GPT-6 Astra rolls out fully, OpenAI claims a Navier-Stokes proof, and more.
Read more →
Qwen3.8-Max-0902 took the Code Arena WebDev crown at 1,691 points for about $5/M tokens vs $20/M for Opus 5. Narrow win, real discount, but the 0902 snapshot is API-only.
Read more →
Alibaba publishes Qwen3.8 weights, Ornith AI claims Opus 4.8 parity with a 397B fine-tune, DeepSeek peak pricing takes effect, and OpenAI cuts GPT-5.6 Sol to $4/$20.
Read more →
Moonshot's 2.8T Kimi K3 launches at $3/$15, Alibaba previews Qwen3.8 Max at WAIC, Thinking Machines ships Inkling, Fable 5 access is extended, and Gemini 3.5 Pro slips.
Read more →
Moonshot's Kimi K3 launched at $3/$15 per million tokens, matching Claude Sonnet's list price. Open weights no longer signal a budget option.
Read more →
Thinking Machines Lab's first model, Inkling, trades top-line leaderboard performance for something harder to buy: an open multimodal base that developers can fine-tune and reshape.
Read more →
Per-token prices say always run max effort on a cheaper tier; per-task costs say it depends. A two-step rule for choosing among GPT-5.6's ten tier-and-effort configs.
Read more →
DeepSeek's shift from price cuts to peak-hour surge pricing reveals capacity constraints. What this means for inference economics and the 'race to zero.'
Read more →
Claude Sonnet 5 lowers token prices but raises cost per task, OpenAI gates the GPT-5.6 preview, Meituan open-sources LongCat 2.0, Fable 5 returns, and DeepSeek plans peak pricing.
Read more →
Tencent Hy3 enters LMRank near DeepSeek V4 Pro with dramatically lower input pricing, strong agentic benchmarks, Apache 2.0 weights, and a clear role as a cheaper production worker model.
Read more →
Meituan's LongCat-2.0 is a 1.6T-parameter MoE coding model trained entirely on domestic Chinese ASICs. It challenges the assumption that frontier AI requires Nvidia hardware.
Read more →
This week in AI: OpenAI launches GPT-5.6 Sol under government gating, Claude Fable 5 stays offline, and Sakana Fugu Ultra hits frontier scores via routing.
Read more →
OpenAI's GPT-5.6 Sol/Terra/Luna shipped only after US government preclearance - a tiered, gated release that could set the template for frontier AI launches.
Read more →
Model routers are here: OpenRouter Fusion's parallel synthesis vs Sakana AI's learned Fugu Ultra orchestration - how they compare on latency, control, and cost.
Read more →
Moonshot AI's Kimi K2.7 Code packs 1T parameters and a $0.95/M-token price, but every launch benchmark is proprietary. Here's why that matters.
Read more →
Claude Fable 5 launches and is shut down by a US export directive within 72 hours, Google releases the DiffusionGemma text-diffusion model, and Moonshot ships Kimi K2.7 Code.
Read more →
Anthropic launched Claude Fable 5 on June 9; a Commerce Department export-control directive forced a worldwide shutdown on June 12. What happened and what it establishes.
Read more →
Anthropic says Claude wrote over 80% of its merged code by May 2026 and task horizons double every four months. What the report shows, and what it actually proposes about pausing.
Read more →
MiniMax M3 pairs 1M-token context with native image and video input at $0.30/$1.20, with open weights promised. Benchmarks, caveats, and how it compares to DeepSeek V4 Pro.
Read more →
Mistral Medium 3.5 is a 128B dense open-weight model with 256K context and 77.6% on SWE-bench Verified. Pricey via API, compelling if you must self-host on four GPUs.
Read more →
DeepSeek locks V4 Pro at $0.435/$0.87 per million tokens, 11x below GPT-5.5 on input. What it costs, why the price may not hold, and how to use it safely.
Read more →
Qwen3.7 Max lists at $2.50/$7.50 with 1M context and near-Opus 4.6 benchmarks, but it writes four times the median tokens. Cheap per token is not cheap per task.
Read more →
DeepSeek V4-Flash brings 1M-token context to open weights: 284B total, 13B active, MIT license, $0.14/$0.28 at launch, and coding scores within two points of V4-Pro.
Read more →
OpenRouter's switcher data shows GPT-5.5's doubled list price hits short prompts hardest (+92%) while long prompts rise about 50%. How to route traffic by prompt length.
Read more →
Every model gets the same detailed brief and builds one HTML file. How the LMRank Showcase works, what the three briefs test, and why we dropped the single score.
Read more →