Latest updates from the world of artificial intelligence
Mistral's new open-weights safety classifier takes the moderation policy as a plain-language prompt at inference time — no retraining — and matches guard models seven times its size while running on a single 16 GB GPU.
Mistral AI
Read more →A hidden entry in Microsoft’s MAI Playground reveals MAI-Realtime, a full-duplex speech model that listens and speaks simultaneously across seventeen languages, with two configurable ways of deciding whose turn it is. Microsoft has confirmed nothing — but the component it would displace inside Azure’s Voice Live service is OpenAI’s GPT-Realtime.
TestingCatalog
Read more →The 2.4-trillion-parameter flagship went broadly available today through Alibaba Cloud's APIs and the new QwenWork beta, with open weights promised for next week. The parameter count is the least interesting part: this is Alibaba reversing course after keeping recent flagships closed. The performance claim behind it — second only to Claude Fable 5 — still has no independent benchmark, and would require a 14-point jump in one generation.
South China Morning Post
Read more →An internal version of Astra resolved ten open problems across group theory, high-dimensional geometry, coding theory, quantum complexity, lattice cryptography and extremal combinatorics, OpenAI says — with the proofs formalised in Lean as machine-checkable certificates. That is a genuine methodological upgrade over May's Erdős result, which rested on expert sign-off. It also leaves the one question a proof checker cannot answer.
The Decoder
Read more →DeepSeek V4 Flash 0731 scores 50 on the Artificial Analysis Intelligence Index, up from 40 in April, with identical architecture, the same 13B active parameters and the same price list — the gain came entirely from post-training. It now sits one point behind GPT-5.6 Luna at roughly 60% lower cost per task, with a 98% cache discount. The nuance most coverage will skip: accuracy was unchanged, and the jump came mainly from a reduced hallucination rate.
Artificial Analysis
Read more →GPT-5.6 Luna drops to $0.20 / $1.20 per million tokens from $1 / $6, and Terra to $2 / $12 from $2.50 / $15. Sol is unchanged, and its new Fast option costs twice as much for about 2.5× the speed. OpenAI credits serving efficiency — but a roughly 20% cost improvement explains Terra's cut and comes nowhere near explaining Luna's, which leaves the gap between the cheapest and best model five times wider than it was a month ago.
OpenAI
Read more →OpenAI reported that enabling retained reasoning and compaction in its Responses API took GPT-5.6 Sol from a 13.3% baseline to 38.3% on ARC-AGI-3's public set, using roughly six times fewer output tokens — a figure above Claude Opus 5's 30.2% record. ARC Prize says its official scores use a standardised setup with no provider-specific settings, and François Chollet allows general-purpose API settings only if clearly reported. Both sides have a point, and the comparison still isn't apples to apples.
OpenAI
Read more →Moonshot AI published Kimi K3's full weights on Hugging Face at 00:00 UTC on July 27, exactly as promised — 2.8 trillion parameters under Apache 2.0, the biggest open-weight release in history. It lands three days after 50 companies signed a letter defending open models, and days after the White House accused Moonshot of distilling an Anthropic model to build it.
Moonshot AI
Read more →OpenAI split GPT-5.6 into Sol, Terra and Luna. Anthropic shipped one model with an effort dial. Put both on the same cost-independent index and the settings line up exactly — including one result that should change how you configure your API calls.
BitsMinds Analysis
Read more →Opus 5 costs the same as the Opus 4.8 it replaces, beats the double-priced Fable 5 on Anthropic's own coding charts, and burns a fraction of the tokens doing it. We lined up every published number — specs, prices, benchmarks — and flagged exactly where the comparisons don't hold.
BitsMinds Analysis
Read more →Anthropic launched Claude Opus 5 on July 24, replacing Opus 4.8 as its workhorse model: a 1-million-token context window, a new 'xhigh' reasoning-effort level above today's max, and unchanged $5/$25-per-million-token pricing. The 'Honeycomb' leak, the Vertex AI sightings, and the specs rumor all turned out accurate — Fable 5 remains Anthropic's top-tier model.
Anthropic
Read more →On July 21 Google released Gemini 3.6 Flash, 3.5 Flash-Lite, and a security-tuned 3.5 Flash Cyber — useful, cheap, efficient. But the flagship Gemini 3.5 Pro is still nowhere to be seen, months past its June target, and Google says it has already begun its most ambitious pretraining run yet: Gemini 4.
The internet is convinced Claude Opus 5 ships any day now — a model codenamed 'Honeycomb' surfaced briefly in Cursor and a Google Vertex AI listing, with leaked talk of a 1M-token context and an 'xhigh' reasoning mode. Here's the honest version: what's actually been spotted, what's pure speculation, and why the timing rumor won't die.
Leak Reports
Read more →China's Moonshot AI released Kimi K3, a 2.8-trillion-parameter open-weight mixture-of-experts model — the biggest ever — that took the #1 spot on Arena.ai's Frontend Code leaderboard at 1,679 points, ahead of Claude Fable 5. It has a 1M-token context, activates just 16 of 896 experts per token, and full weights are promised on Hugging Face by July 27.
Moonshot AI
Read more →Grok 4.5, trained alongside Cursor, doesn't top the leaderboards — Fable 5 and GPT-5.5 still edge it on raw coding benchmarks. But it resolves SWE-Bench Pro tasks in about 15,954 output tokens versus ~67,020 for Opus 4.8, is served at 80 tokens/sec, and costs just $2/$6 per million tokens. xAI's pitch: Opus-class results at a fraction of the time and cost.
xAI
Read more →PrismML has open-sourced Bonsai 27B, a 1-bit build of Qwen3.6-27B that shrinks a model needing ~54GB at full precision down to 3.9GB — small enough to run on an iPhone at 11 tokens/sec while keeping more than 90% of full-precision performance. It's multimodal, handles a 262K-token context, and ships under Apache 2.0.
PrismML
Read more →Thinking Machines Lab's first broadly available model is open. Inkling is a 975B-parameter mixture-of-experts model (41B active), trained on 45T tokens of text, image, audio, and video, with a 1M-token context — released under Apache 2.0 and tuned on the lab's Tinker platform.
Thinking Machines Lab
Read more →ByteDance's new flagship image model reasons about prompts, edits with point-and-lasso precision, writes on-image text in 14 languages, and outputs native 2K — at $0.075 an image. Observers already put its quality at GPT Image 2 level.
TestingCatalog
Read more →
Anthropic pushed free Fable 5 access to July 19 — its third extension in 18 days — as GPT-5.6 undercuts it on price and Grok 4.5 goes live. An Opus 5 leak in Cursor fuels the theory that the extensions are a bridge to the next flagship.
Anthropic
Read more →
OpenAI shipped GPT-5.6 on July 9 as a three-tier family — Sol, Terra, Luna — after a US-government delay. Sol tops TerminalBench 2.1 at 91.9% and is ~54% more token-efficient on agentic coding, though Anthropic's Fable 5 still leads SWE-Bench Pro.
OpenAI
Read more →Meituan open-sourced LongCat-2.0 on June 30 — a 1.6-trillion-parameter mixture-of-experts coding model trained and served end to end on more than 50,000 Chinese-made chips, with no Nvidia hardware in the loop. After two months quietly topping OpenRouter under the alias “Owl Alpha,” the MIT-licensed model claims 59.5% on SWE-bench Pro at a fraction of GPT-5.5’s price.
VentureBeat
Read more →Google let its self-imposed June deadline for Gemini 3.5 Pro pass without a launch, and now says the frontier model will reach general availability in July. The company points to token-efficiency and long-horizon agentic tuning — even as a wave of senior DeepMind researchers heads for the exits.
TechTimes
Read more →Anthropic released Claude Sonnet 5 on June 30 — its most agentic mid-tier model yet, scoring 63.2% on SWE-bench Pro and matching Opus 4.8 on some knowledge work, with introductory pricing of $2/$10 per million tokens. It is now the default model on Claude’s Free and Pro plans.
Anthropic
Read more →DeepReinforce has open-sourced Ornith-1.0, an MIT-licensed family of coding models (9B to 397B) that learn to write their own agent scaffolds during reinforcement training. The 397B flagship resolves 82.4% of SWE-Bench Verified issues — second only to Claude Opus 4.8 — while the 9B model runs on a single 80GB GPU, putting frontier-class agentic coding within reach of self-hosters.
MarkTechPost
Read more →OpenAI released GPT-5.6 on June 27 as a three-model family — the flagship Sol, the balanced Terra and the fast, cheap Luna — but only as a limited preview to about 20 U.S. government-approved partners. Sol leads agentic-coding benchmarks (91.9% on Terminal-Bench 2.1) and matches prior security results using roughly a third of the output tokens, with general availability promised in the coming weeks.
OpenAI
Read more →ByteDance launched its Doubao Seed 2.1 Pro and cheaper Turbo models on June 24, built for the “coding and agent era.” Seed 2.1 Pro scored 1539 on the Code Arena: Frontend leaderboard — eighth in the world and level with Claude Opus 4.6 — while ByteDance claims three core capabilities are comparable to GPT-5.5.
AIbase
Read more →Reports this week floated Claude Sonnet 5 (codename 'Fennec') for the week of June 23, alongside GPT-5.6 — but Anthropic has confirmed nothing. And the same 'Sonnet 5' leak in February shipped as Sonnet 4.6, a reminder to read codenames with caution.
Crypto Briefing
Read more →The Tokyo lab made Fugu generally available on June 22 — a foundation model trained to command a pool of other AIs through one OpenAI-compatible API, pitched as a hedge against single-vendor lock-in.
Sakana AI
Read more →GLM-5.2 is the strongest open-weight coding model yet: MIT-licensed, 744B MoE, 1M context. It beats GPT-5.5 on SWE-bench Pro and trails Opus 4.8 by a point on FrontierSWE — at a fraction of the cost. Our review, with benchmark charts.
The Decoder
Read more →GPT-5.6 is here: OpenAI shipped it on July 9, 2026 as a three-tier family — Sol, Terra and Luna — after a rare US-government delay, with Sol topping several agentic-coding charts. Here's the confirmed launch plus the full pre-launch rumor trail that got us here.
WaveSpeed Blog
Read more →Xiaomi's reasoning model scores 49 on the Artificial Analysis Intelligence Index — #10 worldwide, between Kimi K2.5 and GLM-5 — while running its whole benchmark suite for just $348.
Artificial Analysis
Read more →No OpenAI announcement exists yet, but developer leaks and prediction markets now point to a GPT-5.6 launch in late June — Polymarket prices the June 22–28 window near 83% — with rumored upgrades including a ~1.5M-token context and stronger agentic coding. Treat it all as unconfirmed.
Cryptopolitan
Read more →Zhipu (Z.ai) has launched GLM-5.2, a 744B-parameter Mixture-of-Experts model (40B active) with a 1M-token context window, built for agentic coding and released under a permissive MIT license. The catch: no benchmarks were published at launch, so performance claims remain vendor assertions for now.
Z.ai
Read more →
The Beijing lab’s fifth K2 release in a year is a coding-first, 1-trillion-parameter open-weight model under a modified MIT license — but independent testers say the headline benchmarks are Moonshot’s own and don’t all hold up.
Crypto Briefing
Read more →We pasted identical prompts into Claude Opus 4.8 and Claude Fable 5 — an animated highway interchange, a one-file Space Invaders, a pure-SVG aquarium, and a 3D rocket launch — one shot each, no retries. Everything runs live inside the article. Four builds in, it is dead level at 2–2 — and the launch round flips the very pattern the first three revealed.
BitsMinds Lab
Read more →Anthropic's Claude Fable 5 is the most capable model the public can use today — topping SWE-Bench Pro and excelling at vision and long tasks. It is also the priciest major model and ships with hard safety guardrails. Our first-look verdict.
BitsMinds Reviews
Read more →Anthropic released Claude Fable 5 — the first publicly available model in its powerful Mythos class — scoring 80.3 on SWE-Bench Pro (vs 69.2 for Opus 4.8), priced at $10/$50 per million tokens, with classifier guardrails that block cyber, bio, and distillation requests and fall back to Opus 4.8.
Anthropic
Read more →Prediction markets put it near 94%, a tech reporter says it lands June 9, and internal checkpoints named claude-fable-5 have surfaced — but Anthropic has confirmed nothing. Here is what the noise around Claude Fable 5, the rumored public release of its Mythos model, actually amounts to.
Reporting: Alex Heath; Polymarket
Read more →
MiMo-V2.5-Pro-UltraSpeed broke the 1,000 tokens-per-second barrier on a one-trillion-parameter model using a single eight-GPU commodity node — roughly 15x faster than GPT-5.5 or Claude Opus, and without a line of custom silicon.
MarkTechPost
Read more →The Chinese lab says its new open-weight model pairs frontier coding, a 1M-token context window and native multimodality on a sparse-attention architecture that cuts long-context compute 20x — but the weights and technical report are still days away, so every number is company-reported.
The Decoder
Read more →A model identifier, claude-oceanus-v1-p, surfaced in Anthropic's console on June 3 and went to vetted red-team testers — then, within hours, an unknown actor allegedly resold access through a Chinese API proxy. None of it is officially confirmed, but it offers an early look at the next model in the Mythos line.
Cyber Security News
Read more →On June 3, two image-AI startups shipped models built on the same wager: that the future of generation is precise layout control, not ever-longer prose prompts. Reve's proprietary 2.0 chases 4K and "images you can touch"; Ideogram's open-weight 4 nails in-image text and bounding-box placement. Both are gunning for OpenAI's GPT Image 2 and Google's Gemini.
Latent Space
Read more →OpenAI's life-sciences model just got a major upgrade. Rebuilt on GPT-5.5, it now outscores the general model on a new benchmark called LifeSciBench across drug discovery, genomics, and medicinal chemistry — and OpenAI is widening access beyond launch partners like Moderna, Amgen, and Novo Nordisk.
OpenAI
Read more →At Build 2026, Microsoft’s AI Superintelligence Team unveiled MAI-Thinking-1, its first in-house reasoning model — a sparse Mixture-of-Experts design with 35B active parameters (~1T total) and a 256K context window. Microsoft says human raters prefer it to Claude Sonnet 4.6 and that it matches Opus 4.6 on the SWE-Bench Pro coding benchmark, and it pointedly trained the model from scratch with zero distillation on commercially licensed data — no OpenAI involved.
Microsoft AI / TechTimes
Read more →JetBrains released Mellum2 under Apache 2.0 on June 2 — a 12B MoE model that activates just 2.5B parameters per token for 2x-faster inference, pitched as a fast specialized component for multi-model AI pipelines.
JetBrains AI Blog
Read more →
According to The Information, Microsoft plans to debut a homegrown coding model for GitHub Copilot at its Build conference on June 2–3, part of a wider suite of in-house MAI models led by Mustafa Suleyman. The push to reduce reliance on OpenAI, Anthropic and Google comes as Copilot has lost ground to Anthropic’s Claude Code among developers.
The Information / TestingCatalog
Read more →A BitsMinds analysis. We put each lab’s strongest model through the benchmark portfolio that actually separates frontier systems in 2026 — GPQA Diamond, ARC-AGI-2, AIME, SWE-bench, BFCL and more — then weigh price, context and speed. The short version: Gemini 3.1 Ultra is the sharpest reasoner and the best value, Claude Opus 4.8 owns real coding and agentic work, and GPT-5.5 is the strong all-rounder that everyone already has. Here is the full scorecard.
BitsMinds Analysis
Read more →Anthropic released Claude Opus 4.8 on May 28, 41 days after 4.7 — adding Dynamic Workflows that plan and run hundreds of parallel subagents per session, a user-facing effort dial, a 2.5× faster fast mode at 3× lower cost, and its “most honest” self-review behavior yet, at 4.7’s price.
Anthropic / TechCrunch
Read more →Inside the 512K-line Claude Code source map that npm shipped on March 31 was a feature literally called Undercover Mode — and inside Undercover Mode was a hard-coded list of names it was supposed to scrub from public builds. That list spells out Opus 4.8, Sonnet 4.8, a never-released Sonnet 4.7 and animal codenames Capybara and Tengu — with Capybara described internally as a new tier sitting above Opus.
WaveSpeed AI / ChaoBro / AI Nexus Daily
Read more →
Qwen3.7-Max-Preview ships with a 1M-token context, extended-thinking mode, and benchmark gains across CritPt, Humanity's Last Exam, and Terminal-Bench Hard — clearing Gemini 3.5 Flash on the AA Intelligence Index.
Qwen (Alibaba Cloud)
Read more →
Test UI strings inside Gemini reveal a new Google video model that lets users remix and edit clips directly in chat, with a likely debut at I/O 2026 on May 19-20.
9to5Google
Read more →
Mira Murati's Thinking Machines Lab released a research preview of TML-Interaction-Small, a 276B mixture-of-experts model that responds in under 0.4 seconds and listens while it speaks — outpacing GPT-realtime-2 and Gemini 3.1 Flash Live on the FD-bench latency test.
TechCrunch
Read more →OpenAI rolled out three new voice models on May 7 — a reasoning agent with a 128K context window, a 70-language live translator at $0.034 a minute, and a streaming Whisper that transcribes as you speak.
OpenAI
Read more →
Gemini 3.1 Flash-Lite is now generally available on Vertex AI and Gemini Enterprise, pitched as the cheapest, fastest member of the Gemini 3 family for high-volume agentic, classification, and tool-use workloads.
Google Cloud Blog
Read more →
Mistral fuses chat, reasoning, and coding into a single 128B dense model and pairs it with Vibe — async cloud-based coding agents that hand back work as pull requests instead of terminal output.
Mistral AI
Read more →
Chinese lab DeepSeek released two open-weight V4 preview models with 1M-token context windows, matching closed-source rivals on reasoning while undercutting them on price.
TechCrunch
Read more →NVIDIA's new 30B-A3B mixture-of-experts model unifies vision, audio, and language into a single open system, delivering up to nine times the throughput of comparable open omni models for AI agents.
NVIDIA Blog
Read more →
OpenAI's GPT-5.5, codenamed "Spud," brings 82.7% Terminal-Bench 2.0 performance and autonomous multi-step task execution — a step toward OpenAI's unified AI super app.
TechCrunch
Read more →DeepSeek's V4-Pro and V4-Flash arrive with 1.6 trillion parameters, a 1M-token context window, and Apache 2.0 licensing — matching near-frontier performance at 21x less cost than Claude Opus 4.7.
Simon Willison
Read more →
Google's Gemini 3.1 Pro hits a verified 77.1% on the ARC-AGI-2 reasoning benchmark — more than double its predecessor — and arrives with a 2-million-token context window, multi-step agentic coding, and support across Gemini API, Vertex AI, and NotebookLM.
Meta Superintelligence Labs unveils Muse Spark, a fast multimodal model that deploys parallel AI subagents, analyzes medical images, generates functional websites, and surfaces social-powered shopping recommendations.
Meta
Read more →xAI has launched Grok 4.3 in beta with native video processing, PDF and spreadsheet generation, and tighter Grok Computer integration — behind a $300/month SuperGrok Heavy paywall ahead of a broader May rollout.
TechSifted
Read more →
Moonshot AI has released Kimi K2.6, a 1-trillion-parameter open-source model capable of coordinating 300 parallel sub-agents across 4,000 steps — available on Hugging Face under a Modified MIT License.
MarkTechPost
Read more →Google DeepMind released Gemma 4 on April 2, a family of four open-weight multimodal models under Apache 2.0 that bring frontier-level reasoning and agentic capabilities to phones, edge devices, and developer environments.
Google DeepMind
Read more →
Anthropic has revealed Claude Mythos 5, the first publicly acknowledged 10-trillion-parameter AI model, engineered for cybersecurity, advanced coding, and deep academic reasoning — but not yet available to the public.
Anthropic / AI Analytics Diaries
Read more →
Microsoft unveils MAI-Transcribe-1, MAI-Voice-1, and MAI-Image-2 through its Foundry platform, marking the company's boldest push yet to build AI infrastructure independent of its OpenAI partnership.
TechCrunch
Read more →
Claude Opus 4.7 is now generally available, bringing a 13% coding benchmark improvement, 3x higher image resolution support, and a new xhigh effort setting -- at the same price as Opus 4.6.
Anthropic
Read more →
NVIDIA unveils Ising, the world's first family of open-source AI models built to accelerate fault-tolerant quantum computing through AI-powered calibration and error correction.
NVIDIA Newsroom
Read more →OpenAI releases GPT-6 with a 40% performance leap, a 2 million token context window, and a unified super-app merging ChatGPT, Codex, and the Atlas browser into a single agent experience.
OpenAI
Read more →Google's flagship Gemini 3.1 Ultra achieves a verified 94.3% on GPQA Diamond and 77.1% on ARC-AGI-2, setting new state-of-the-art benchmarks for complex scientific and logical reasoning.
Google DeepMind
Read more →OpenAI's new small models are designed to function as parallel subagents inside larger AI workflows — fast, cheap, and capable enough to handle the bulk of agentic work.
OpenAI
Read more →
NVIDIA debuted the Nemotron 3 family at GTC 2026, releasing Nano immediately and previewing Super (49B) and Ultra (253B) variants, along with three trillion tokens of pre-training data.
NVIDIA Newsroom
Read more →
Meta CEO Mark Zuckerberg unveiled Llama 5, a 600-billion-parameter open-weight model trained on 500,000 Blackwell GPUs that Meta says outperforms GPT-5 and Gemini on key benchmarks.
FinancialContent / Meta AI
Read more →Google launches Gemma 4, a family of open models ranging from 2B to 31B parameters that deliver unprecedented intelligence-per-parameter, outperforming models 20 times their size on key benchmarks.
Google Blog
Read more →Meta debuts Muse Spark, a proprietary multimodal model built by Alexandr Wang's elite superintelligence team, marking a dramatic shift away from the company's open-source Llama strategy.
TechCrunch
Read more →Anthropic has begun limited enterprise testing of Claude Mythos 5, a 10-trillion-parameter model that the company calls a 'step change' in AI performance, excelling at cybersecurity, complex coding, and academic reasoning.
Fortune
Read more →Google's Gemini 3.1 Pro sets a new industry benchmark with a 2 million token context window and 77.1% on ARC-AGI-2, enabling AI to reason over entire codebases and lengthy documents in one pass.
MarkTechPost
Read more →Google unveils Gemini 2.0 with advanced multimodal capabilities and deep integration across all Google services.
Google AI Blog
Read more →
Anthropic launches Claude 4 with dramatic improvements in text comprehension, coding, and mathematical reasoning.
Anthropic
Read more →OpenAI has announced GPT-5, its most advanced model yet, with enhanced reasoning capabilities and multimodal features.
OpenAI Blog
Read more →