|
|
MODEL
SEISMIC
2026-07-24
Claude Opus 5 — Anthropic's new Opus nears Fable 5 at half the price
Anthropic's new Opus tier comes close to Fable 5 while holding the Opus 4.8 price.
What is it?
Claude Opus 5 is Anthropic's newest Opus-class model, live today on Claude.ai, the Claude API, Claude Code, and Claude Cowork. It lands close to the frontier intelligence of Claude Fable 5 at half the price, with the same $5/$25 per 1M token pricing as Opus 4.8.
How does it work?
The claude-opus-5 checkpoint roughly doubles Opus 4.8 on Frontier-Bench v0.1 and scores 3× the next-best model on ARC-AGI 3. A Fast mode is available at 2× price for higher throughput.
Why does it matter?
Opus-tier pricing has stayed flat while capability effectively doubled — teams already on Opus 4.8 get a major step up with zero billing change. On OSWorld 2.0 Opus 5 surpasses Fable 5 at about a third of the cost.
Who is it for?
Claude API customers, agent builders, and Claude Max and Pro users — Opus 5 is now the default model on Claude Max.
|
|
|
|
TOOL
MAJOR
2026-07-24
Claude Code 2.1.219 — Opus 5 becomes default, subagents nest to depth 3
Anthropic's terminal coding agent adopts Opus 5 as its default and lets subagents spin up their own subagents three layers deep.
What is it?
Claude Code 2.1.219 makes Opus 5 the default Opus model in the CLI, adds a 1M-token context in fast mode, and enforces a strict sandbox network allowlist that refuses non-allowed hosts without prompting.
How does it work?
Subagents can now spawn their own subagents down to three levels deep — the previous hard cap was one. Set CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH=1 to restore the flat behavior in workflows that assume a single fan-out.
Why does it matter?
Users get Opus 5's stronger long-horizon coding without changing any flags, and the deeper subagent tree lets one workflow decompose tasks into workers-of-workers. The stricter sandbox allowlist removes approval-fatigue in headless CI runs.
Who is it for?
Claude Code power users, teams running Claude Code in CI, and anyone building multi-agent workflows on top of the CLI.
|
|
|
|
ECOSYSTEM
MAJOR
2026-07-24
DeepSeek retires deepseek-chat and deepseek-reasoner — API aliases die today
DeepSeek's legacy API aliases go dark at 15:59 UTC — migrate apps to V4-Flash or V4-Pro before the cutoff.
What is it?
DeepSeek retires its legacy deepseek-chat and deepseek-reasoner API aliases at 2026-07-24 15:59 UTC. After the cutoff, any request that names either model returns an error — no grace period, no automatic fallback.
How does it work?
The two aliases were compatibility shims routing to the same underlying V4-Flash model in different modes. Callers must now name the target directly: deepseek-v4-flash (non-thinking, $0.14/$0.28 per 1M) or deepseek-v4-pro (thinking, $0.435/$0.87 per 1M).
Why does it matter?
Every app, script, gateway config, or eval suite still using the old names will 4xx the moment the cutoff passes. A straight model-string swap to deepseek-v4-flash does not raise your bill.
Who is it for?
Any developer using the DeepSeek API — especially teams running scheduled jobs or eval pipelines that haven't been touched since the aliases were announced.
|
|
|
|
TOOL
MAJOR
2026-07-23
ChatGPT Voice lands on desktop — OpenAI puts GPT-Live in the Mac and Windows apps
ChatGPT Voice, powered by OpenAI's new GPT-Live model, is now inside the Mac and Windows apps.
What is it?
ChatGPT Voice lands in the Mac and Windows desktop apps (v26.715), powered by GPT-Live — OpenAI's new speech-to-speech model built for low-latency turn-taking. All paid plans (Plus, Pro, Business, Edu, Enterprise) get it globally.
How does it work?
GPT-Live is a single speech-in, speech-out model — no stitched TTS stages — and plugs into the same Codex, Computer Use, and local-file tools already in the desktop app. The Mac-only Appshots mode lets the assistant see the active window while you talk.
Why does it matter?
Voice used to be a phone feature, so anyone coding or reviewing spreadsheets had to swap devices. Bringing GPT-Live to the desktop keeps the conversation on the same machine where the work lives — Appshots turns it into a screen-aware pair programmer without copy-paste.
Who is it for?
Paid ChatGPT users on Mac or Windows who want a hands-free assistant while coding, writing, or reviewing on-screen work.
|
|
|
|
TOOL
MAJOR
2026-07-23
AMD Helios — 72-GPU rack-scale AI system aimed straight at Nvidia's NVL72
AMD's first rack-scale answer to Nvidia's NVL72: 72 MI455X GPUs, 31 TB of HBM4, and hyperscaler commitments already signed.
What is it?
AMD Helios is a rack-scale AI system combining 72 Instinct MI455X GPUs, 18 sixth-gen EPYC Venice CPUs, 31 TB of shared HBM4 memory, and AMD Pensando networking — delivering 2.9 exaflops of FP4 inference. Microsoft, OpenAI, Meta, Oracle, and Anthropic are named launch customers.
How does it work?
Helios treats the rack as a single accelerator: the 72 MI455X GPUs share all 31 TB of HBM4 through a high-bandwidth fabric, so one training or inference job sees the entire pool. AMD ships ROCm, Pensando data-plane, and pre-integrated firmware as one bundle.
Why does it matter?
Nvidia's GB200 NVL72 has been the only rack-scale option for frontier training — the shortage drives GPU prices up. Helios is the first credible alternative with signed hyperscaler commitments, finally giving buyers a second source.
Who is it for?
Hyperscalers, AI labs, and sovereign cloud operators planning the next generation of frontier training clusters.
|
|
|
|
MODEL
MAJOR
2026-07-23
FLUX 3 x mimic — the FLUX 3 backbone drives factory robots at 101 ms
A FLUX 3 spin-off that lets one on-robot GPU pilot a real factory arm in near real time.
What is it?
FLUX 3 × mimic is a joint video-action model from Black Forest Labs and mimic robotics. It reuses the FLUX 3 video backbone and adds a custom action decoder so one network handles both perception and robot control — tested on Audi factory lines at 101 ms end-to-end reaction time.
How does it work?
The action decoder attends to FLUX 3's latent video stream and outputs joint commands in overlapping chunks while the next prediction is already running. Mimic quantizes the model to run inference in under 80 ms on a single NVIDIA RTX 5090 sitting on the robot itself.
Why does it matter?
FLUX 3 × mimic clears 95% success on soft-body kitting tasks with up to 10× fewer demonstrations than typical VLA models, and it runs on the robot — no datacenter inference, closer to plug-and-play factory deployment.
Who is it for?
Industrial robotics teams and embodied-AI researchers looking for a path to practical factory automation without massive sim-to-real pipelines.
|
|
|
All releases at ai-tldr.dev
Simple explanations • No jargon • Updated daily
|
|
