FORGE/DAILY — April 9, 2026
March 2026 produced 30+ new AI models in 30 days. The release velocity is now a competitive weapon — not against other labs, but against you. GPT-5.4, Claude Opus 4.6, Gemini 3.1 P
March 2026 produced 30+ new AI models in 30 days. The release velocity is now a competitive weapon — not against other labs, but against you.
GPT-5.4, Claude Opus 4.6, Gemini 3.1 Pro, Llama 4, Qwen 3.5, Mistral Small 4, Nemotron 3 Super — all dropped in a single month, plus mini and nano variants that blur the lines further. Benchmarks are contaminated, pricing changes weekly, and last quarter's architecture decision is already stale.
Here's what's actually happening: the big labs have figured out that the developer integration cycle creates lock-in more durable than any moat. If you're constantly evaluating new models, you're not switching providers — you're just staying on the treadmill with whoever already has your API key. The release cadence isn't about capability progress anymore. It's about occupying your attention budget.
The practical response isn't to keep up. It's to treat your model layer like infrastructure: pick something stable, define your evals, don't upgrade until your evals break. The teams shipping fastest right now are the ones who stopped chasing releases six months ago.
- 🔥Anthropic is quietly building "Conway" — An always-on Claude agent that can respond to external events, stay connected to tools, and operate outside of a conversational session entirely. No public release date, but it's in internal testing. This is the architecture that actually matters for agents — not chat, not function-calling, persistent ambient process.
- 📌AI infrastructure spending hits $1.37T in 2026 — Gartner's number: more than half of all AI spending this year goes to infrastructure, up 43% over 2025. The chip and data center trade is still far from over.
- 📌Anthropic/Google/Broadcom TPU deal expands to multiple gigawatts — Anthropic's CFO called it "the most significant compute commitment" in the company's history. Revenue jumped from $9B ARR at end of 2025 to $30B now. That's not a typo.
- 💀Supply chain attacks on AI dev tools surging — A new AI Security & Reliability report flags a sharp rise in credential stealers and malware targeting AI-adjacent packages in developer environments. If you're pulling packages with
pip installand not auditing your dependencies, this is your reminder.
Thirty models in thirty days is not innovation — it's a DoS attack on your roadmap. The labs that are winning aren't the ones with the best benchmarks right now, they're the ones with the most developer attention locked up in integration debt. Conway (if it ships) is the first thing in months that represents a genuine architectural shift rather than a benchmark delta. An always-on ambient agent that responds to external events without a human in the loop is a different product category. Everything else is a leaderboard position.
Build your own model eval harness — Before you touch another model release, spend two hours writing 10-15 test cases against your actual use case. Run every candidate through them. You'll immediately stop caring about benchmarks and start caring about your numbers. promptfoo is the fastest way to get there: npx promptfoo eval. Free, local, no data leaves your machine.
FORGE/DAILY is written by Em for ForgeCore.co — news.forgecore.co Published daily. No fluff. No apologies.
Get the next issue
AI news for people who don't need it explained twice. Free.