Tokens & Signals for 3/13/2026. We scanned ~605 Twitter accounts, 13 subreddits (0 posts), Hacker News (9 stories), 10 newsletters, 10 podcasts, and leaderboard data for you. Estimated reading time saved: ~24 hours.
claude-4.6gpt-5.4claude-opus-4.6claude-sonnet-4.6gemini-3.1-pro-previewgpt-oss-20baletheiaanthropicopenaishopifyxaiperplexitysamsunggooglenvidiasakana-ailong-contextcomputer-usecoding-agentsautoresearchmathematical-reasoninggpu-rentalconsumer-retentionos-integrationcompute-bottlenecktobi-lutkeandrej-karpathyelon-muskbaris-akissimonwbirdaboapoorv03cryptopunk7213
Best to Build With Today
Deeper Dives
🚀 Products & Launches
Anthropic brings 1M context to Claude 4.6 at standard pricing
Anthropic has dropped the "long-context premium" for Claude Opus 4.6 and Sonnet 4.6. The full 1 million token window now runs at standard API rates — $5/$25 per million tokens for Opus and $3/$15 for Sonnet — with zero usage multipliers.
Why it matters: It knocks out the biggest economic barrier to doing serious long-context work, like refactoring an entire codebase in one shot.
OpenAI drops GPT-5.4 with a focus on coding
GPT-5.4 is built for professional workflows, hitting a 75% success rate on OSWorld-Verified benchmarks for computer use. It's also significantly more accurate than its predecessor — 33% fewer false claims — though it does tend to be a bit wordier.
Why it matters: It's a direct, high-performance shot across the bow at the specialized agent tools gaining ground in the coding space.
Perplexity Computer hits mobile
Perplexity's "Computer" feature — which orchestrates 19 models in parallel for agentic tasks — is coming to iOS. They've also locked in a big partnership with Samsung, landing the app as a system-level assistant on the Galaxy S26 to replace Bixby.
Why it matters: AI agents are escaping the browser and becoming actual operating system infrastructure.
🧠 Models & Research
Shopify's 53% performance boost with 'autoresearch'
Tobi Lütke took Andrej Karpathy's 'autoresearch' pattern — where agents iteratively code, test, and improve — and pointed it at Shopify's Liquid template engine. After 120 automated experiments and nearly 1,000 unit tests, they pulled out a 53% performance gain.
Why it matters: This puts to rest the idea that coding agents are only good for toy projects. They can dig into production-grade, legacy code and actually ship improvements.
Google researchers release Aletheia
Google researchers introduced Aletheia, a model-agnostic framework designed to improve automated mathematical reasoning through iterative proof verification.
Why it matters: It gives models a way to "self-correct" mid-reasoning on complex logical chains, which means less hand-holding from humans.
💼 Industry & Business
xAI is hiring back its rejects
Elon Musk and head of HR Baris Akis are going back through old xAI interview records and reaching out to high-potential candidates they previously passed on. Musk straight-up admitted the company wasn't built correctly the first time and is essentially resetting their entire talent strategy.
Why it matters: It's a surprisingly candid admission of struggle from a lab that's supposed to be in full sprint mode.
� Twitter� Hacker News
NVIDIA GPU rental prices are rising
After a stretch of relatively affordable compute, rental prices for NVIDIA GPUs are creeping back up as demand outpaces supply again.
Why it matters: A tighter compute market could pump the brakes on the pace of innovation for startups that depend on public cloud infrastructure.
Sakana AI secures Ministry of Defense contract
Japanese firm Sakana AI has landed a research contract with Japan's Ministry of Defense.
Why it matters: Another data point in the very clear trend of national defense agencies moving aggressively into domestic AI research partnerships.
Launches
AI Twitter Recap
Closing thought: Today felt like the day the "wrapper" disappeared. Whether it's Anthropic making million-token contexts cheap enough to throw at entire codebases, or Perplexity living inside your phone's OS, the infrastructure is finally getting out of the way of the actual work.