Saturday, July 25, 2026

Anthropic shipped Claude Opus 5 — near-Fable-5 intelligence at half the price, with a new low/medium/high effort dial and a fresh #1 on SWE-benchA benchmark that tests whether a model can resolve real GitHub issues by editing code in popular Python repositories.Jimenez et al. 2023 ↗ Verified. DeepSeek V4 went GA around WAIC and its legacy API aliases retired yesterday, so migrations are now mandatory. And OpenAI had its fourth outage in four days this morning, taking ChatGPT, the API, and Codex down worldwide before recovering by midday ET.

TLDR

Anthropic releases Claude Opus 5 — frontier-class coding/agents at flat $5/$25 pricing, with a low/medium/high effort toggle to trade cost for capability (Jul 24) Trending Pricing
DeepSeek V4 hits GA around WAIC; legacy deepseek-chat/deepseek-reasoner aliases retired — introduces first-ever peak/off-peak API pricing (Jul 24) Pricing
OpenAI's 4th outage in 4 days — ChatGPT, API, and Codex down globally this morning, recovered by ~noon ET (Jul 25) Production
Travis Kalanick's Atoms raises $1.7B in an a16z-led round to build physical/industrial AI (Jul 23) Trending
Humanoid becomes Europe's first pure-play humanoid-robotics unicorn at a $1.35B valuation on a $152M Series A (Jul 21) Robotics/IoT

Action Items

Things your engineering team should look at or do this week.

Confirm any DeepSeek integrations moved off the legacy aliases. deepseek-chat and deepseek-reasoner were fully retired after Jul 24 15:59 UTC and now map to deepseek-v4-flash. Update model IDs to deepseek-v4-flash/deepseek-v4-pro, and if you batch jobs, schedule them off-peak to dodge the new peak-hour surcharge (~2x during Beijing 9am–12pm / 2pm–6pm). (DeepSeek API docs | TechNode)
Trial Claude Opus 5's effort dial for cost control before re-baselining agent workloads. Pricing is unchanged from Opus 4.8 ($5/$25 per M tokensA small unit of data, such as part of a word, that AI models read and generate. Many AI services price by tokens in and out.NVIDIA ↗), but the low/medium/high toggle lets you cut reasoning depth (and tokens) on simple prompts. Given it now leads SWE-bench Verified, re-run your coding-agent evals against it. (Anthropic | VentureBeat)

Models & Tools

Anthropic launches Claude Opus 5 (Jul 24) Pricing Trending — Anthropic's fourth model in under two months (after Mythos 5, Fable 5, Sonnet 5). Positioned as "close to the frontier intelligence of Claude Fable 5 at half the price," with a 1M-token context windowHow much text a model can see and use at once while writing a reply.Anthropic ↗, an xhigh reasoning mode, and a user-facing low/medium/high effort toggle to balance cost vs. capability. Now the default on Claude Max and top model on Pro; pricing held flat at $5/$25 per M tokens. (Anthropic | Fortune)
Alibaba releases Qwen-Audio-3.0-TTS (Plus and Flash) (Jul 20) Open Source — New text-to-speech models in the Qwen-Audio line, giving self-hosters a fresh open option for voice/agent pipelines. Worth a look if you're building voice AI and want to avoid per-call TTS API costs. (llm-stats)

APIs & Pricing

DeepSeek V4 reaches GA around WAIC; legacy aliases retired (Jul 24) Pricing — V4-Pro (1.6T total / 49B active) and V4-Flash (284B / 13B active) both ship with 1M-token default context, 384K max output, JSON/tool-calls, and thinking + non-thinking modes. The deepseek-chat and deepseek-reasoner aliases were fully retired after Jul 24 and now resolve to deepseek-v4-flash. DeepSeek is also rolling out its first peak/off-peak surcharge pricing. (Reported anticipated in yesterday's digest — now confirmed GA.) (DeepSeek API docs | TechNode)

Industry & Business

Travis Kalanick's Atoms raises $1.7B (Jul 23) Trending Robotics/IoT — The Uber co-founder's physical-AI startup closed a $1.7B equity round led by a16z (Ben Horowitz joins the board), with Bain Capital, Uber, Fifth Wall and others participating. Atoms targets industrial automation across food, mining, and transportation using software, sensors, and robotics — a sign capital is still flooding into "bits-to-atoms" physical AI. (SiliconANGLE | PYMNTS)
Humanoid becomes Europe's first pure-play humanoid-robotics unicorn (Jul 21) Robotics/IoT — The UK industrial-robotics startup raised a $152M Series A at a $1.35B post-money valuation (led by Prime Movers Lab; Bosch, Schaeffler among investors), taking total funding to $270M. Beta robots roll out to logistics/manufacturing/retail customers in Q4 2026; partners include SAP, Nvidia, Bosch, and Siemens. (Forbes | Yahoo Finance)
Tempus AI acquires Personalis for ~$1.5B (Jul 19) — The precision-medicine/AI-diagnostics player consolidates genomic data assets, continuing 2026's pattern of AI-driven M&A concentrating in data infrastructure. (Intellizence)

AI in Production

OpenAI hit by global outage — its fourth in four days (Jul 25) Production — Around 5am ET, ChatGPT, the developer API, and Codex all logged elevated error rates worldwide; users couldn't load conversations, reach chat history, or run prompts, and downstream apps depending on the endpoints failed with "too many concurrent requests." Services fully recovered by ~11:57am ET. The recurring disruptions this week are a reminder to build fallbacks/timeouts around single-provider dependencies. (BleepingComputer | The Next Web)

Benchmarks & Evals

SWE-bench Verified reshuffles at the top (Jul 24) — With Opus 5's launch, the coding leaderboard now reads: Claude Opus 5 at ~97.0%, GPT-5.6 Sol at ~96.2%, and Claude Fable 5 at ~95.0%. The frontier is tightly clustered — differences are now small enough that cost, latency, and your own eval harness matter more than the headline number. (llm-stats SWE-bench Verified)

Strix — open-source AI penetration testing (usestrix/strix) — Multi-agent tool that autonomously runs recon, exploitation, and post-exploitation to find and validate app vulnerabilities with real proof-of-conceptProof of conceptPublic exploit code that shows a vulnerability can actually be triggered.FIRST ↗ exploits (not just false positives), with CI/CD integration. ~39K stars and climbing fast — worth evaluating as an automated security-testing layer in your pipeline. (GitHub)

A quieter, model-heavy day after the week's security drama.