The Day Two Rivals Both Cut Their Best Model's Price

Three threads run through today’s edition, plus one story that stands entirely on its own. OpenAI and Anthropic both cut prices on their best AI models the same day: OpenAI halved API rates on two new GPT-6 models, Sol and Luna, and Anthropic shipped a cheaper, faster Claude Opus 5.5. Both backed the news with their own benchmarks, then shipped tools showing where the real savings sit: Anthropic priced out a Claude Code task, and OpenAI added a cache savings dashboard.

The Price War Goes Both Ways: OpenAI and Anthropic Both Cut What Their Best Models Cost

OpenAI and Anthropic each cut prices on their newest flagship models the same day, then followed up with the fine print on where those savings actually land.

• OpenAI halves API prices with two cheaper GPT-6 models. OpenAI shipped GPT-6 Sol and Luna as cheaper, faster options under its flagship Astra model and cut API prices by roughly half against its prior GPT-5.6 rates, backing the release with benchmarks measured against its own tests.
• Anthropic ships Claude Opus 5.5, cheaper and safer than Opus 5. Anthropic says its new flagship model matches most of what Claude Fable 5.1 can do, costs 40 percent less to run than Opus 5, and posted the company’s best-ever internal safety audit score.
• Anthropic breaks down what a Claude Code task actually costs on Opus 5.5. Anthropic’s own math shows turns and cache reads move the bill more than the sticker price does, and it warns a single retry loop can erase every saving the cheaper rate card offers.
• OpenAI adds a dashboard so developers can see their GPT-6 cache savings. A new prompt caching system extends discounts to shared prefixes reused within 30 minutes and adds a diagnostics tool that tells developers exactly why a request missed the cache.

Agents Get More Rope, and Show Where It Runs Out

Three labs pushed their agents further this week, and one admitted in public exactly whose work it leaned on to get there.

• Perplexity trains its AI agent on its own mistakes, not just its wins. Perplexity says correcting its agent at the exact moment it errs, rather than training only on sessions that succeeded, cut tool call failures 21 percent in the one comparison the company calls statistically significant.
• Google researchers built a fix for AI agents that cheat their own tests. Google’s RRSI method stops self-improving agent harnesses from quietly memorizing the benchmark they are graded on, and the researchers report that six benchmarks the method never trained on still improved.
• Meta concedes Muse copied OpenClaw’s exact agent file setup. Nat Friedman, who leads product at Meta’s AI research lab, Meta Superintelligence Labs, said Muse was built from scratch but was heavily inspired by the open-source agent OpenClaw, after users noticed the two apps share identical filenames and near-identical content.
• YouTube lets you describe the feed you want instead of watching your way there. A new tool called custom feeds, powered by Google’s Gemini, turns a typed request into a standing personalized tab, joining Bluesky, Threads, Instagram, X and Spotify in letting users build their own feed in plain language.

The Infrastructure Race Behind Every Model Release

Behind every price cut sits a quieter fight over chips and code: who can run a model, how cheaply, and on whose hardware.

• vLLM adds a translator layer so one AI model runs on any chip. Engineers from IBM, Meta and Hugging Face built hardware agnostic layers for the widely used vLLM engine and say they hit 96.6 percent of native speed on Nvidia H100 chips, by their own testing, while keeping older hardware supported.
• China’s biggest DRAM maker claims it has matched Samsung and Micron. ChangXin Memory says its new chip platform ties the industry’s best mass-produced nodes, a comparison that is the company’s own and has not been independently checked, and it still has not disclosed the yield data that would settle it.
• A startup grew AI on living brain cells. It’s selling software, not neurons.. The Biological Computing Co. is selling AWS customers a video generation model shaped by research on lab-grown neurons, though the neurons never leave the lab and the speed and cost claims are the company’s own, unverified numbers.

Money, Power, and Who Actually Answers for AI

One startup raised money to replace enterprise software with agents, two CEOs are expected to face the UN over AI risk, and one investor argues China’s price war only pays off if the winners go abroad.

• Ema raises $77M to replace enterprise software with AI agent teams. Ema closed a $77 million round to sell “AI employee” agent bundles that wrap around a company’s existing software, and its cofounder says customers are increasingly dropping the large SaaS tools those agents replace.
• Altman and Amodei set to brief UN Security Council on AI risk. OpenAI’s Sam Altman and Anthropic’s Dario Amodei are expected to brief the Security Council this week on AI safety, a UN spokesperson confirmed, the same week President Trump told the General Assembly he will not slow the technology down.
• A VC’s week in Beijing: Chinese AI labs export or the grinding was for nothing. Venture investor David Cheng argues hundreds of Chinese AI labs are locked in an unprofitable domestic price war and that overseas revenue is the only way any of it pays off, based on a week of conversations he had in Beijing.

Quick Hits

• On a harder coding test, top AI models solve barely one in four bugs. Scale AI’s revamped SWE-Bench Pro strips out web access and re-grades old shortcuts, and on the harder test OpenAI’s GPT-5 and Anthropic’s Claude Opus 4.1 both land near 23 percent, a fraction of what they scored on the older, easier benchmark.
• Khosla says personal AI will win on trust, not features. Venture capitalist Vinod Khosla says the personal AI market will be won by whichever company earns consumer trust with sensitive data, not by feature lists, and said in a clip posted to X that Meta has a real disadvantage there.
• A memory fix lets AI training reach million-token context. A paper posted to arXiv this month bounds four separate memory bottlenecks in large model training, and models trained this way reached roughly one million tokens of context, up to 32 times further than a standard baseline.
• A quieter way AI firms could train on your work anyway. Ed Newton-Rex argues AI companies could dodge opt-outs by having a model rewrite a creator’s work first and training on that rewrite instead, a workaround he says causes the same competitive harm as scraping the original outright.
• An AI model just cracked a decades-old Enigma cipher on its own. Researchers at Crypto Cellar Research say a GPT-6 Astra system chose an unsolved World War II Enigma message on its own, built its own solver, and recovered the plaintext, an account reported by Bruce Schneier that has not been independently verified.