Latest updates from the world of artificial intelligence
Z.ai’s new open-weight model scores 57.0 on the Artificial Analysis Intelligence Index — six points behind Claude Opus 5 and under two behind GPT-5.6 Sol, at $0.15 per million input tokens. It loses the flagship fight and wins the one that matters for volume: it beats Sonnet 5, Terra and Luna outright.
Artificial Analysis
Read more →The stealth model that appeared on OpenRouter under the provider name “stealth” has a name at last. Z.ai’s GLM-5.3-Flash is a 320-billion-parameter multimodal mixture-of-experts model with a million-token context window and MIT-licensed weights — reportedly, a preview served entirely on Chinese-made accelerators.
MarkTechPost
Read more →IBM published three open-weight models — 3B, 8B and 30B — with a switchable thinking mode across the family and reinforcement learning inside real software-engineering, terminal and web-search sandboxes for the two larger sizes. The 30B scores 57.0 on SWE-Bench Verified, and the whole family runs on hardware you own.
The Decoder
Read more →A reasoning model with a million-token context window appeared on OpenRouter on August 20 under the provider name "stealth". Five days later, no lab has admitted building it.
TechCrunch
Read more →Google shipped Gemini 3.7 Flash just three weeks after 3.6 Flash, claiming a jump from 34.4% to 43.6% on FrontierCode and from 49% to 65.3% on DeepSWE — and priced the introductory tier at half what 3.6 Flash cost.
Z.ai shipped GLM-5.3 on August 14 without retraining its base model. Every gain comes from scaled post-training — and on the company’s own coding benchmark it edges past Claude Opus 4.8 while spending roughly 60% fewer output tokens.
MarkTechPost
Read more →
We gave Claude Opus 5, Fable 5 and Sonnet 5 the same three build briefs — an animated SVG fairground, a falling-sand physics sandbox and a self-solving Rubik’s cube — one run each, no retries, no browser. All nine builds run live inside the article. Opus 5 takes it 8–6–3, and the biggest surprise is the clock: the largest model was the fastest on every round, by a factor of nearly three.
BitsMinds Lab
Read more →No blog post, no changelog, no press release — just an edit to the API pricing page. DeepSeek-V4-Pro-0813 is now generally available at $0.87 per million output tokens, with benchmark gains that nobody outside DeepSeek has reproduced.
OpenRouter
Read more →Artificial Analysis scored Grok 4.6 at 61 on its Intelligence Index — level with GPT-5.6 Sol at max, behind Fable 5 at 62 and Opus 5 at 63. It costs $2/$6 per million tokens against $5/$25 and $5/$30, and finishes long agentic tasks in ~53 turns where Opus 5 takes ~103. It also trails both rivals by eight points on the hardest software-engineering evals.
Artificial Analysis
Read more →OpenAI split its Daybreak security program into Blue and Red tiers and handed the Red tier a purpose-trained model that clears 95% of advanced cyber tasks — against 1.5% for the same frontier model with its guardrails intact.
The Decoder
Read more →Meta released Muse Glimmer under Apache 2.0 — a 30-billion-parameter agentic model distilled from Muse Spark that quantises under 20 GB and runs on a single consumer GPU.
Meta AI Research
Read more →OpenAI slowed work on its next frontier model after internal evaluations suggested Astra may cross the Critical cybersecurity threshold — a level no model has triggered before. Weights are locked down, some internal work is paused, and outside evaluators are being called in.
TechCrunch
Read more →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.