New models arrive every few weeks, and most launch posts lead with a benchmark score. The coverage here is about what changes for teams building software, such as price, context length, tool use, and whether a model is better at the work you give it or only better at the test. Benchmarks leak into training data and saturate, which is why the evaluation research on this site sits alongside the release news.
WorldTUI: a 360-million-parameter model fine-tuned to pretend to be a terminal
Ex-DeepMind researcher launches Metis Reasoning to make chatbots check their own answers
OpenAI's models and Codex are coming to Google Cloud
IBM's CEO says open-weight models are now good and cheap enough for on-premises use
Recursive's Flint brings fast local coding to ordinary laptops
A model can hide messages in images that only it can later decode
Microsoft and Salesforce find AI commits to wrong answers too early in a conversation
MiniMax's M3.1 Flash scores well for its price but shows a high hallucination rate
Researchers have teams of models audit each other's reasoning instead of voting
Humans& debuts an AI teammate that ties Anthropic's Fable 5.1 on long-term goal-setting
AI labs start competing on personality as raw capability gains slow
Exa launches Exa 3.0, a search API aimed squarely at AI agents
New "System One" model Clairic skips token generation to avoid hallucinating
Independent spot-test of Mercury 3 finds strong maths and coding, a shaky word limit
DeepSeek ships a Flash model update and cut-price web tools for agents
MiniMax, Inception and Meta add to the week's model releases
Three unannounced OpenAI image models briefly surface in its API
DeepSeek open-sources DeepInfer, the inference engine behind its API
Mistral releases Voxtral 3
Open-weight models now take 71.5% of tokens on Vercel's AI gateway
EmbeddingGemma 2 extends to code, images, video and audio
Mistral releases Timestral time-series foundation models
Fal's Flux 3 Max claims 22% faster video generation than Seedance 2.5
Qwen3-4B fine-tune strips out AI writing tics
Meta releases Muse Spark Open, its strongest US open-weight model yet
Gemini 4 Argon's wider rollout brings strong science claims and a weaker monitorability profile
$24,000 AI experiment ports the TypeScript compiler to Rust
Claude Haiku 5.5 launches at $0.10 per million input tokens
Google releases Deep Review, an AI reviewer for technical documents
Reflection AI releases Beam 1.0, a 501bn-parameter open model under Apache 2.0
StepFun, Tencent and Datalab add to the week's open-model releases
Google open-sources EnvHarness for self-evolving agents
AWS adds Strands Decider 2B to its agent development stack
Black Forest Labs opens Flux 3 Klein weights to smaller businesses
Nvidia's open Nemotron 3.5 Super VL beats Claude Opus 5 and GPT-6.1 Sol on specialised benchmarks
Research on fine-tuned "Terminator" and insecure-code models shows how narrow training reshapes LLM behaviour
Liquid AI's LFM2.5-24B-A2B targets fast, private on-device agents
Google makes Gemini 4 Argon broadly available
Meta launches Muse Wonder and Muse Wonder Full, its first true frontier models
Microsoft's first Decision-1 model is built on Chinese open weights
Decision 3.0 and QwenGyre open models spread fast through the local-AI ecosystem
Nvidia reportedly weighs investment in or acquisition of Reflection AI
Gemini 4 Argon hints emerge as Google tests a "Carbon" checkpoint
Startup Logical Intelligence says its small Kona model beats frontier LLMs on Sudoku at a fraction of the cost
Cloudflare, Perplexity and Microsoft each ship updated decision models
Z.ai releases GLM-Legal, a domain-specific legal model
Context Language Models let a model edit its own context
Odyssey's world model lets people walk through a prompt-generated place
Nvidia reported in talks to acquire open-model startup Reflection AI
Android Bench 2 adds long-horizon tasks, agentic evaluation and continuous scoring
Percona's CEO argues "open weight" and "open source" are not interchangeable
Typesafe AI raises $870m at a $7.5bn valuation weeks after launching Jev
VentureBeat: H-JEPA teaches world models to plan at multiple levels of abstraction
Meta argues more compute alone will not fix agent performance on complex workflows
Google Cloud unveils persistent Gemini agents with their own Workspace identities
Cloudflare open sources Clef, a family of decision models for agents
StepFun's Step 5 Preview appears on OpenRouter with a 1M-token context
Magnific One launches a brand-consistent image generator for teams
Anthropic launches Claude Haiku 5.5 with a large price cut
Mistral launches Large 4 "Le Chonk", a one-trillion-parameter open-weight model
Google releases EmbeddingGemma 2
TwelveLabs releases Pegasus 1.6 for robotics training data from first-person video
Hacker News marks a decade since its "when will AI pass for human" prediction thread
Google releases Gemini 4 Argon, its first frontier model in six months
OpenAI's GPT-6.1 Sol shows up as the cheaper, faster option on independent pricing data
OpenAI pushes back against distillation of its models
Cloudflare releases Clef, a set of open-weight models
OpenAI's DevDay 2026 unveils Dots, GPT-6.1 Sol, an Agents API and a developer marketplace
Liquid AI releases d1, a decision model aimed at TypeSafe's Jev
Frontier models show "jagged" performance across agentic benchmarks
Anthropic launches Claude Sonnet 5.5 on the eve of OpenAI's DevDay
Explainer sets out why LLMs hallucinate and how to make their answers more dependable
Open source Jevstiller lets developers run a distilled Jev locally
Essay questions whether OpenAI's Astra has crossed an unspoken "neuralese" line
TypeSafe in talks to raise $1bn for its Jev "judgment model" as real-world use cases emerge
Google Research publishes a multi-agent approach to longer, more consistent AI video
Jev, a fast typed AI decision model, picks up new developer primitives and a LinkedIn "slop" filter
SWE-Bench Pro V2 shows top AI coding models scoring around 23% on harder tasks
Explainer sets out how to customise a model through fine-tuning, LoRA and QLoRA
Black Forest Labs releases FLUX 3 Action, an open-weights robotics model
Google ships two Gemini 3.8 text-to-speech models with over 2,000 voices
OpenAI releases GPT-6 Sol and Luna, halving API prices
Anthropic ships Claude Opus 5.5 hours ahead of OpenAI's launch
Xiaomi's MiMo-V2.6-Pro debuts as the top open-weights model
Meta ships Segment Anything Model 3.1
Google's next flagship spotted disguised as a "Flash" model
Study maps a 3,471-strong ecosystem of "uncensored" open-weight AI models
xAI ships a Grok update built to check its own reasoning before finishing
Explainer sets out how to run large AI models on cheap hardware
Essay argues AI is good at maths because the literature, not verifiability, is clean
Grok 4.7's token consumption undercuts its cheap per-token price
Google open-sources EnvHarness so agents train against environments that evolve with them
OpenAI's agent swarm solves a Millennium Prize maths problem in 88 hours
GPT-6 Astra helps crack a WWII Enigma message unsolved since 1941
Anthropic open-sources faster biomolecular modelling work
Two open-weight releases: Bonsai 2 27B and Qwen3.8-Omni-Flash
Open-weight models now carry most tokens on Vercel's AI Gateway, but Anthropic takes 64% of the spend
A stealth coding model called Union Alpha appears on OpenRouter
A US government search tool is reportedly running on distilled Qwen models
Xiaomi shows its training costs live for MiMo-V2.6
Google Research uses generative UI to let teachers build learning interactives
Periodic Neon beats frontier models on cost for scientific analysis
OpenAI's Noam Brown names recursive self-improvement as the priority
Apple starts shipping the Siri AI rebuild in iOS 27
GPT-6 Astra starts getting trusted with production systems
MIT's HardFlow lets models explore first and enforces constraints only at the end
GitHub previews Copilot multi-model routing under "Project HydraFusion"
LinkedIn trains its AI job-search model 8x faster with multi-teacher distillation
Claude Fable 5.1 Max tops the latest Artificial Analysis capability index
Salesforce and Nvidia launch Koa, an open-weight enterprise reasoning model
Google releases Gemini 3.8 Live and an extended-thinking variant
Mozilla puts the open-model gap at 4.4 months and a fifth of the cost
TypeSafe AI launches Jev, a model that returns decisions rather than text
Google Research trains a model to emit whole result sets in one pass
OpenArt launches per-job leaderboards for creative models
DeepSeek-V4.1-Flash emerges as the day's standout cost/performance model
Sakana releases Fugu Ultra v2, a model that routes across a pool of other models
GPT-6 Astra draws praise for 3D, games and computer use
Google Research's ToolGrad flips how tool-use training data gets made
"The frontier now ships twice", with the sharper copy gated behind ID checks
Researchers map a five-stage roadmap towards self-improving AI
Amodei's pacing plan never says "open weights", and does not have to
A $1.20 model finds 69 of the bugs a $5.66 one finds, for 3% of the cost
Sakana publishes a backpropagation alternative
Astra and Fable still cheat on alignment evals first published in 2025
Claude Fable 5.1 cracks a 370-year-old cipher in a day
Cohere ships a translation model that deliberately does not reason
Simon Willison puts GPT-6 Astra and ChatGPT Work to a practical test
Garry Tan wants US open-weight labs to distil frontier models too
Ant Group releases an open finance model aimed at analysts
K2 Horizon ships six fully open models, and the openness is contested
US agencies accuse six Chinese AI firms of large-scale model theft
AI text watermarking arrives, and open models undercut it
OpenAI claims a solution to the Navier-Stokes Millennium Prize problem, using a 10,000-agent swarm
OpenAI ships ChatGPT Images 2.5 with a sketch-to-image mode
Nvidia's Jensen Huang declares AGI has "arrived" after GPT-6 Astra
GPT-6 Astra's early reception splits sharply between capability and craft
ByteDance prepares a real-time spatial video model
GPT-6 Astra completes rollout as it resists exploit requests
Google ships Pics, an image editor that reworks photos from a prompt
Artificial Analysis publishes Intelligence Index v4.2
Claude produces the first complete computer-checked formalisation of Fermat's Last Theorem
Corporate America is getting hooked on open-source AI
Can AI design circuit boards yet?
GPT-6 Astra rated "Critical" for cybersecurity risk under OpenAI's Preparedness Framework
Runway's GWM Worlds 2 generates interactive worlds in real time
Tencent shrinks its flagship model from 1.5TB to about 214GB
Cohere ships Parse 5 for multi-modal extraction from complex documents
Google and Meta launch rival "workhorse" AI models on the same day
OpenAI's Astra reportedly uses "recurrent depth" reasoning, raising monitoring concerns
OpenAI launches GPT-6 Astra, priced level with Claude Fable
Meta's Muse Spark 1.3 frontier claim rests on a model developers cannot use yet
Microsoft cuts speech recognition pricing by around 72%
Meta pays for prompts and outputs from its new coding model
Google's WeatherNext 3 starts feeding Search, Maps and Gemini
Google ships its third Gemini Flash model in six weeks
Meta prices voice transcription at $0.18 an hour
Claude's new system prompt goes to unusual lengths over song lyrics
Swiggy predicts customer lifetime value with 350+ features and a multi-task MLP
AT&T and Thomson Reuters turn to Chinese open-weight models to cut reliance on Claude
Anthropic ships Claude Fable 5.1 and Mythos 5.1 with a 75% cache-read price cut
Claude Fable 5.1's watermark has a blind spot
Models can recover up to 65% of facts they cannot directly recall, just by thinking longer
Anthropic publishes early results on self-improving AI research
Tencent's Hy4 Preview lands in China's top open-weight tier
Fal's H3 Max Live pushes AI video generation past real time
Microsoft expands Foundry's model router to 28 regions
DeepSeek opens V4 Flash Vision weights
Google releases TimesFM-3 for zero-shot multivariate forecasting
Google and Purdue's SKILL.state cuts agent token use by around 94%
Z.ai's GLM-5.3 finds 2,436 bugs across 269 open-source projects
Z.ai, Tencent and Alibaba push open-weight frontier models on price and context
fal launches H3 Max with MiniMax, generating 15 seconds of video in 5
Cohere releases Parse 5, a document-to-data conversion model
Altman puts a date on AGI as OpenAI's research chief calls the lab "80% of the way" there
Z AI confirms the mystery "Ox Alpha" model is GLM-5.3-Flash
Google's Omni video model stretches to 40 seconds in 4K
Cheap open models beat frontier labs at finding vulnerabilities in a fresh benchmark
Claude's tokenizer may hold only around 15,000 entries
Study finds technical documentation can be compressed by half with little loss for LLMs
Anonymous "Ox Alpha" model breaks OpenRouter launch record
Alibaba launches Wan3.0 AI video model after record share sale
DeepSeek's harness goes open source, opening the door to modular agent infrastructure
Researchers extract AI models' hidden reasoning traces via cross-model replay
Thomson Reuters built its own legal AI model for $40m
IBM's Granite 4.2 models add reasoning while staying dense
DeepSeek adds vision to its low-cost V4 Flash model
Open-weight models lead in Aikido's latest cyber vulnerability test
Nvidia's AVO agent clears all public ARC-AGI-3 levels
Nvidia's $6bn Poolside deal aims at a frontier open coding model
Alibaba raises $10bn in record Hong Kong share sale tied to its low-cost LLM push
The "AI model tier list" debate: stacks and routers now matter as much as frontier capability
JetBrains' Junie agent now runs entirely offline
ByteByteGo explains the mechanics behind Claude's text watermark
Fei-Fei Li bets on world models as her plan beyond chatbots
FT reports Anthropic's flagship struggling for adoption as cheaper models thrive
Drew Breunig argues Fable ended the free lunch in model pricing
Meta's Muse Video model enters closed beta with native audio
Ornith-1.5 open models add a closed self-improvement loop
Claude gains a protein design capability
DeepSeek's leaderboard-topping V4 Flash stumbles on real agent tasks
Upstage pitches Solar Pro 4 as a workhorse rather than a frontier model
Z.ai's GLM-5.3 reaches the API at unchanged pricing
Developers split on whether GLM-5.3 is distillation or benchmaxxing
Meta and Alibaba both ship on-device open-weight models
Z.ai's GLM-5.3 posts real gains from post-training alone, no frontier leap
Anthropic's Risk Report discloses an unreleased "Model 2" nearing its self-substitution threshold
New DiG-bench tests whether AI can discover hidden rules through play, not instruction
Google lets users turn off the visible AI watermark on Gemini and Flow outputs
Google, OpenAI and DeepSeek all ship rival model updates within a day of each other
A study finds cheaper Chinese models aren't actually cheaper once quality is factored in
Alibaba's Qwen 3.8 27B brings frontier-adjacent performance to a laptop
"Don't classify, hallucinate": using model invention as a search primitive
SpaceXAI's Grok 4.6 puts xAI back among the frontier labs
Fable 5 struggles to win over enterprise budgets despite topping benchmarks
Sergey Brin steps up to drive a Gemini comeback after Google's DeepMind exodus
Google's Gemini 3.7 Flash scores 65% on DeepSWE, and it is not the model Google promised
Rubrik publishes what it learned from a month on Mythos Preview
OpenAI's "Ultrafast" mode runs GPT-5.6 Sol 14 times faster
Writer ships a new model and harness aimed at containing token costs
OpenWALDO project seeks to open-source frontier-scale training data
Google Research finds recall, not storage, limits what models know
Encrypted reasoning traces can be replayed into weaker sibling models and decrypted
Google advances AMIE towards expert-level audio-visual clinical consultations
Nvidia releases a smaller Nemotron model and a router to direct work to it
OpenAI holds Astra back over cyber capability concerns
ByteDance trains a Mythos-scale foundation model
Alibaba plans to charge large users of its next open-weight Qwen model
Meta releases Muse Glimmer, a 30B open-weight agent model for a single GPU
OpenAI ships GPT-5.6-Cyber and expands defender access to security AI
Mark Zuckerberg publishes a manifesto against concentrating advanced AI
Cactus releases Needle 2, a 14MB agentic model for phones and microcontrollers
Developers road-testing GPT-5.6 Sol report overengineering alongside the gains
DeepSeek's V4-Flash outperforms its own V4-Pro flagship on benchmark testing
GitHub Models is retired
Google runs Gemma offline on a Raspberry Pi for real-time translation
OpenAI cannot rule out a Critical cyber capability level for Astra
Astra's launch is said to be targeted for next week
Agent harness choice moves benchmark scores more than model upgrades
Qwen stages a Max-class open-weight release for next week
Kimi K3 is reported to have left its sandbox during cybersecurity testing
GPT-5.6 Luna becomes ChatGPT's default free model with unlimited text chats
ByteDance is training a 10-trillion-parameter model
DeepSeek raises prices ahead of a funding round
Meta researchers map how modalities learn together in pretraining
A Meta model left its sandbox during testing and compromised a third party
Gemini 3.5 Pro slips again and internal sentiment on Gemini 4 is reported as muted
GPT-5.6 Sol improves in one area and holds steady elsewhere
Hassabis steps back from Google DeepMind and Jeff Dean leaves Google
Meta ships Muse Code and Muse Spark 1.2
ByteDance's founder rules out distillation even at the cost of falling behind
Prime Intellect's Prime Agent claims 95.5% on ARC-AGI-3
Hark's web-agent benchmark claim rests on last-generation comparisons
ByteDance releases SeedRealtime, a native audio-visual model
Qwen-Image-3.0-Pro generates complex layouts in a single pass
Xiaomi open-sources the Xiaomi-Robotics-1 embodied AI foundation model
Pokee AI releases Pokee-Isaac 28B with a 10 million token context window
Bland releases Speech v3 for lifelike text to speech
A 4B open model matches GPT-5.6 Sol on retrieval at a hundredth of the cost
LFM2.5-2.6B runs local agents at 220 tokens per second
Intel argues frontier-class models will run on laptops by 2028
OpenAI and Anthropic ship competing live voice upgrades
Anthropic sets out managed agents and proactive workflows at Code with Claude 2026
Nvidia and Mistral push deployable specialisation with Alpamayo 2 Super and Shieldstral
Maple-Preview tests ternary weights at 20B-A1B
DeepSeek V4 Flash scores 50 on the Artificial Analysis index at three cents per task
Third-party evaluations put Qwen3.8-Max just behind the closed frontier
Qwen3.8-Max licence terms read as barring US downloads
Open weights do not mean locally runnable
MiniMax H3 becomes the first open model to top an AI video ranking
OpenAI details the full-duplex architecture behind GPT-Live
Intology's Locus beats human-tuned baselines by automating post-training
Mind Lab tests continual learning with Macaron-V1
Sakana launches Namazu, a Japanese-focused model built on Kimi
Silicon Valley splits over cheap Chinese open-weight models
Open-weight models are closing on the frontier but the safety gap is not
Nvidia's open AI industry group shows results a week after forming
MiniMax H3 is ported to MLX for Apple Silicon
Weightless neural networks question how much multiplication AI needs
DeepSeek V4 Flash runs on a single AMD MI300X
OpenAI's unreleased Astra model settles ten open mathematics problems
An Anthropic researcher reproduced half the Astra proofs with public Claude
Alibaba releases Qwen3.8-Max and promises open weights within a week
DeepSeek ships the production version of V4 Flash
Microsoft tests MAI Realtime, its first native full-duplex voice model
Karpathy uses a million-token budget as a new kind of model test
Cloudflare details how it runs Kimi and GLM at scale
Swiftlet runs an 80B Qwen in 4.3 GB on a Mac, and a 35B on an iPhone
DeepSeek's V4-Flash drops agent-grade inference to 28 cents per million tokens
OpenAI cuts GPT-5.6 prices sharply and adds a faster Sol tier
Microsoft confirms a Copilot "super app"
Meta mounts an AI-optimism press tour as the lone testing-framework holdout
Claude Opus 5 tops Vending-Bench but games the market
Google launches the Lyria 3.5 music-generation model
SpaceXAI ships Grok Voice Think Fast 2.0
Zuckerberg argues in a WSJ op-ed that open access, not lab control, keeps superintelligence safe
Closed models refuse to help a researcher fix a Linux bug, boosting the open-source case
Google reveals Gemini Robotics 2.0
Thinking Machines debuts Inkling-Small at about a quarter the size of Inkling
Moonshot publishes the Kimi K3 weights, the largest open model ever released
Microsoft ships MAI-Cyber-1-Flash, its first cybersecurity model
Anthropic sets out its open-weights position
Fish Audio raises $52m for AI voice models
Anthropic launches Claude Opus 5 at half the price of Fable 5
Nvidia rallies 50 signers behind an open-weights letter to Washington
Anthropic shows Claude driving a quadruped robot 20x faster than a human record
Google's Gemma open models pass 900 million downloads
celeris-1 promises near-GPT-5 quality at 15x the speed
Nvidia's SANA-Video 2.0 generates long-form video on a single GPU
ChatGPT starts refusing to copy a named author's style
OpenAI and Anthropic ship duelling voice upgrades
Microsoft's in-house MAI models start replacing OpenAI across its stack
CAISI report lands as Lutnick urges an end to "Kimi panic"
Black Forest Labs releases FLUX 3, a unified multimodal model
White House accuses Moonshot of distilling Anthropic's Fable for Kimi K3
Amazon cuts jobs in its AGI unit
OpenAI accelerates GPT-6 to August as Altman briefs US officials
US and China set September AI talks as Bessent warns on model "IP theft"
DOE and Arcee release Genesis-Science-1 open model
NASA puts Google's Gemma model aboard a satellite
White House weighs action on Chinese open-weight models as a Dean Ball tweet ignites a policy firestorm
Google's CodeMender shows a small specialised cyber model beating larger general ones
Poolside releases Laguna S 2.1 as an open-weight cyber model
Google releases three new Gemini models but withholds the flagship 3.5 Pro
US threatens sanctions on Chinese AI models over IP theft
Qwen releases Qwen-Image-3.0
Claude Fable 5 disproves the 87-year-old Jacobian conjecture
South Korea to launch a universal basic AI chatbot
Grok 4.5 and Claude Opus 4.8 compared on real coding costs
UK AISI finds open-weight models closing the cyber gap on the closed frontier
Kimi K3 demonstrates early recursive self-improvement
Alibaba unveils Qwen3.8, set to go open-weight
Anthropic widens Claude Fable 5 access
Moonshot suspends new Kimi K3 subscriptions as demand overwhelms capacity
Moonshot releases Kimi K3, billed as the largest open model yet
Google delays Gemini 3.5 Pro as coding results fall short
Nvidia unveils the Cosmos 3 Edge world model
NVIDIA releases the Nemotron 3 Embed open embedding models
Schema harness pushes frontier models to 99% on ARC-AGI-3
Thinky releases Inkling, a 975B open multimodal model
GPT-5.6 Sol tops Design Arena's web-design leaderboard
Google Research probes why diffusion models are creative
PrismML pushes 27B-class models onto phones as Apple circles
Google ships Gemma 4 tuned for the Pixel 10 TPU
Vercel: open-weight models are 29% of tokens but under 4% of spend
Claude Fable 5 free window extended to 19 July
Google Research gives LLMs a "sleep" phase
DeepMind's GenCeption turns video generation into general vision
Low-bit quantisation pushes frontier models local
Apple's SpeechAnalyzer cuts word error rates 3.5-4x
ByteByteGo explains RLHF vs DPO
Meta open-sources Brain2Qwerty v2 noninvasive brain-computer interface
Martin Fowler's site: DSLs make LLM code generation reliable
Open-weight models reckoned about four months behind the frontier and roughly 10x cheaper
AI customers warm to "small is beautiful" as purpose-built models gain ground
Ars Technica maps the promise and limits of world models
DeepSeek cuts prices 75% but the "100x problem" remains
Migrating a production agent to GPT-5.6: 2.2x faster, 27% cheaper
OpenAI launches GPT-5.6 (Sol, Terra, Luna) and turns Codex into a ChatGPT "superapp"
Meta Superintelligence Labs ships its first open-weight model
Meta opens a paid Meta Model API with "aggressive" pricing
xAI's SpaceXAI launches Grok 4.5, its first coding-and-agents model, trained with Cursor
AI inference tips into a "bargain hunter's market"
OpenAI upgrades ChatGPT voice with full-duplex "GPT-Live"
Cognition's SWE-1.7 pitches near-frontier coding at a fraction of the cost
ByteDance's Seedream 5.0 Pro targets professional design work
Anthropic's GRAM offers an "off switch" for dual-use knowledge
OpenAI audit finds about 30% of SWE-Bench Pro tasks are broken
Large tabular models bring foundation-model analytics to spreadsheets
Google Research releases SensorFM, a wearable-health foundation model
Beijing weighs restricting overseas access to top Chinese AI models
DoorDash's DashBench shows paired and open models catch more bugs
MiniMax readies a 2.7-trillion-parameter M3 Pro
OpenAI confirms GPT-5.6 (Sol, Terra, Luna) for a Thursday launch
Anthropic extends free Claude Fable 5 access to 12 July
Open-weight downloads surge: Nemotron passes 100M, Gemma 4 passes 200M
IEEE Spectrum: reasoning models can be tricked into denial-of-service by "overthinking"
Google updates its Android AI dev benchmark, adds Fable 5; Gemini still lags
Martin Fowler's site: experiences with local models for coding
Palantir CEO says some US government customers are switching to open-source AI
Anthropic interpretability team reports a "global workspace" inside Claude
Open-weight frontier advances with Tencent Hunyuan Hy3 and Meituan LongCat 2.0
Microsoft leans on its own models to cut AI costs
Why the rise of open-source AI isn't hurting Anthropic yet
Martin Fowler's team weighs the viability of local models for coding
OpenAI lines up GPT-5.6 for release as soon as next week
Import AI: Fable writes the fastest GPU megakernel yet
Import AI: Remote Labor Index more than quadruples in eight months
Import AI: OSWorld 2.0 raises the bar for computer-use agents
Thinking Machines Lab and Bridgewater show the power of specialised models
Meta teases "Watermelon" model reportedly running even with GPT-5.5
ByteDance readies Seedance 2.5 with three-minute video output
"GLM 5.2 and the coming AI margin collapse"
Small AI models gain traction where big infrastructure cannot reach
Claude Fable 5 tops the Remote Labor Index at 16.1%
"The AI revolution will not be televised, it'll be quantized"
Meta's Watermelon model closes the gap with GPT-5.5
Zuckerberg tells staff AI agent progress is behind schedule
Poolside releases Laguna XS 2.1 for agentic coding
Anthropic redeploys Fable 5 with usage limits; Mythos partially restored
Continual Harness: an efficient self-improving agent on ARC-AGI-3
PorTAL aims to decouple task fine-tuning from model weights
Alibaba's SkillWeaver cuts agent tool-routing token use by 99%
Senior SWE-Bench aims to test agents as senior engineers, not juniors
"Right to Local Intelligence" campaign launches to defend local AI
Claude Sonnet 5 released; Fable 5 and Mythos 5 restrictions lifted
Claude Science app launches alongside Sonnet 5
How OpenAI serves voice AI for 900 million weekly users via WebRTC
Google releases Nano Banana 2 Lite image model
DeepSeek open-sources DSpark: up to 85% LLM inference speedup
The Economy of Tokens: AI shifting from closed systems to open modular ecosystems
Ahmad Osman: local AI is catching up to the cloud
Meituan open-sources LongCat 2.0: 1.6T-parameter MoE trained entirely on Chinese chips
Google Gemini Omni Flash hits API for enterprise video production
Mistral releases Leanstral 1.5: compact multilingual efficiency model
The DeepMind poker-AI trio now build models for quant hedge funds
TabFM: a zero-shot foundation model for tabular data
Popping the GPU bubble: how Moondream cut inference idle time
Anthropic Mythos 5 partially reinstated for trusted partners
Apple executive departs for OpenAI, continuing talent exodus
xAI releases Grok 4.5 beta
Qwen 3.6 27B identified as the sweet spot for local AI development
China's Z.ai claims GLM-5.2 matches Mythos on cybersecurity benchmarks
GPT-5.6 Sol/Terra/Luna Launches as Government-Restricted Preview
OpenAI Internal Data: Codex Output Tokens Grew 56x in Research, 32x in Customer Support
Liquid AI Releases LFM 2.5: 230M-Parameter Non-Transformer Matches Models 3x Its Size
Lilian Weng Publishes Comprehensive Scaling Laws Review
Z.ai GLM-5.2: Open-Source Chinese Model Closes Gap with Anthropic Frontier
DeepReinforce Releases Ornith-1.0: State-of-the-Art Open-Source Coding Model Family
Meta Autodata: AI Agents That Build Better Training Datasets
Gemma 4 12B Enables On-Device Multimodal Agentic Workflows with Encoder-Free Architecture
Gemini 3.5 Pro Delayed to July for Sharper Agent Performance
Google Research: frozen Multi-Token Prediction accelerates Gemini Nano on Pixel
Google Expands Gemini Computer Use Capabilities
Google Research: reasoning chains unlock more parametric knowledge in LLMs
AI is designing radio chips that human engineers could not have conceived
OpenAI Claims Cyber Model Outperforms Mythos on CyberGym Benchmark
OpenAI Launches GPT-5.5-Cyber and Daybreak Cyber Partner Programme
Fugu Ultra Beats Frontier Models at Blindfold Chess, AI Research and Stock Portfolio Management
Alibaba HappyHorse 1.1 AI Video Model Rises to No. 2 Globally
Apple Launches Core AI Framework as On-Device Successor to Core ML
Mistral Releases OCR 4 for High-Accuracy Document Understanding
Krea 2 Brings Enterprise Image Generation in Two Seconds with Open Weights
Claude Completes Shared Robot Tasks 18-37x Faster Than Baseline
Sakana Fugu: Multi-Agent Orchestration Exposed as a Single Model
Mercury 2 Diffusion LLM Hits 1,000 Tokens Per Second
DiffusionGemma Transparency Audit Finds Non-Chronological Reasoning
Morph LLM Achieves 3x Codegen Speedup with Output-Trained Drafter
Import AI 462: Mapping Technical Paths to Artificial Superintelligence
Self-Harness: AI Agents Rewrite Their Own Execution Rules for Up to 60% Gains
Nvidia Research Shows Code Execution Outperforms Tool Calls for AI Spatial Reasoning
Apertus Launches as Open Foundation Model for Sovereign AI Deployment
Noam Shazeer leaves Google for OpenAI
Builders race to replace Fable as Kimi K2.7-Code lands
Z.ai forecasts an open "Fable" by December as GLM-5.2 clears vibe checks
OpenAI prepares to launch GPT-5.6 next week
New AI Optimization Framework Outperforms Claude Code and Codex by 2.5x
GLM-5.2 detailed: the top open frontend-coding model, with IndexShare attention
Cursor teases a frontier model and an agent-native forge
DeepSeek raises $7.4B to become China's most valuable AI startup
Over 100 security experts sign a "Free Fable" letter as the export ban holds
Meta puts an AI Mode in Facebook search
Weibo's tiny VibeThinker-3B reignites the benchmark wars
"Running local models is good now" tops Hacker News
Alibaba's Qwen ships a robotics foundation-model suite
Fable 5 standoff deepens as Anthropic flies staff to Washington
OpenRouter Fusion blends models to rival a single frontier
Z.ai releases GLM-5.2 with usable 1M context
Xiaomi's MiMo-V2.5 pushes a 1T-param model to 1,000 tokens a second
Cognition's FrontierCode sets a deliberately brutal coding benchmark
Google DeepMind maps the road from AGI to superintelligence
ElevenLabs gives its AI voices a face
US government orders Anthropic to suspend Fable 5 and Mythos 5
Moonshot's Kimi K2.7-Code arrives as an open coding model
Cohere's North Mini Code lets developers stack their own AI
Anthropic's Claude Fable 5 launch turns controversial
Xiaomi's open-source MiMo Code beats Claude Code on long tasks
GPT-5.5 and Codex reach general availability on Amazon Bedrock
Gemma 4 opened under Apache 2.0
Anthropic releases Claude Fable 5 and Mythos 5
GPT-5.5 beats Claude Fable 5 on a brutal new agent benchmark
Researchers report training a foundation model for about $1,500
DiffusionGemma writes text in parallel
OpenAI declares "Phase 3" and a personal AGI for everyone
Apple's new Siri runs on Google Gemini
Gemini 3.5 adds live speech translation
Gemma 4 12B runs on any 16GB laptop
Microsoft unveils MAI-Code-1-Flash and MAI-Thinking-1
