Agentic Business

DISTRIBUTION WARS

OpenAI plugs ChatGPT into OpenClaw. Anthropic slams the door.

via TheNextWeb

OpenAI just turned ChatGPT into the backend for the biggest open-source agent project in history - the same project Anthropic spent April kicking off its lawn.

Subscribers can now point their existing $23/month ChatGPT plan at OpenClaw's 3.2M-user runtime and spin up autonomous agents without a separate API bill. That's the inverse of Anthropic's April call to block Claude on the same platform - a decision Anthropic framed as margin protection while OpenAI frames distribution as the moat.

So you've got two philosophies hardening into competitive lines: OpenAI eating compute cost to win the default-agent slot, Anthropic guarding subscription economics and forcing builders onto first-party surfaces. Both can be right. Only one can be the path of least resistance for an OpenClaw user opening the app tonight.

The split, in one paragraph

• OpenAI: ChatGPT subscription = production-viable OpenClaw backend, no extra metering.
• Anthropic: Claude blocked on OpenClaw as of April; routes builders back to first-party.
• Builders targeting OpenClaw's installed base now have a clear default - and a clear non-option.

Why it matters: If you ship on top of OpenClaw, your model-selection menu just got a lot shorter - and the broader question of who pays for inference at the agent layer just got its first real answer. Read more.

BENCHMARK BREAK

Mistral Medium 3.5 hits 77.6% on SWE-Bench - and ships the cloud agent to use it

via Mistral AI

Mistral just stopped being the polite European option on your model dropdown. Medium 3.5 - a 128B model - posted 77.6% on SWE-Bench Verified, putting it in the same neighborhood as Opus 4.6 and GPT-5.5 on the only coding benchmark anyone still respects.

Bigger story for builders: the model shipped alongside Vibe Remote Agents, a cloud runtime that runs coding tasks asynchronously in isolated sandboxes, pings you when done, and auto-opens pull requests against your GitHub repo. No babysat terminal, no local GPU, no background process eating your laptop battery.

The pivot is the point. Mistral is moving away from "run it on your machine" toward distributed cloud execution as the default agent surface - the same direction Anthropic and OpenAI have been pulling the market for two quarters. The benchmark gets the headline; the runtime is what changes how you actually build.

• 77.6% SWE-Bench Verified - competitive with Opus 4.6 / GPT-5.5 tier.
• Vibe Remote Agents run in isolated cloud sandboxes, no local resources tied up.
• Automatic PR creation against connected GitHub repos when a task completes.

Why it matters: Cloud execution plus auto-PR is the friction-killer agentic coding has needed - and there's now a third lab credibly playing in the top tier. Read more.

METER'S RUNNING

Notion ends the free Custom Agents trial. The credit clock starts today.

via Notion

The free hors d'oeuvres are gone - Notion is rolling out the bill on Custom Agents, six months after the launch party.

The free trial ends today, May 3, with paid pricing via Notion Credits kicking in May 4. To soften the blow, Notion swapped in a fresh model bench - GPT-5.4 Mini/Nano, Haiku 4.5, MiniMax M2.5 - claiming 35–50% cost reductions and up to 10× fewer credits on repetitive tasks.

This is the first real-world price tag on enterprise agent deployment from a mass-market SaaS, and it's a useful tell. Notion isn't competing on raw capability - it's competing on cost-per-task, which is exactly the axis builders should be tracking once the novelty wears off and the finance team starts asking pointed questions.

• Free trial ends May 3; metered Notion Credits pricing begins May 4.
• New model lineup: GPT-5.4 Mini/Nano, Haiku 4.5, MiniMax M2.5.
• Claimed 35–50% cost reductions; up to 10× fewer credits for repetitive flows.

Why it matters: If you're modeling agent unit economics inside a real product, today's Notion price sheet is the cleanest public data point you've had so far. Read more.