Claude now has a smarter way to run agents without blowing your budget.

The advisor strategy lets you pair Opus as an advisor with Sonnet or Haiku doing the actual work, getting you near Opus-level intelligence at a fraction of the price.

The idea is simple. Your cheaper executor model handles the task end-to-end, and only calls in Opus when it hits a genuinely hard decision. Opus weighs in with a plan, and the cheaper model executes. All inside a single API request!

The numbers back it up too. Sonnet with an Opus advisor scored 2.7 percentage points higher on SWE-bench Multilingual than Sonnet flying solo, while actually costing 11.9% less per task.

For Haiku, the gains are even more remarkable, its BrowseComp score more than doubled, from 19.7% to 41.2%, with Opus in its corner.

It's a one-line change to your existing Messages API call. Add the advisor tool, set a max_uses cap so costs don't spiral, and you're done.

OpenAI is reshuffling its subscription tiers to keep up with how heavily people are actually using Codex.

A new $100/month Pro plan is joining the lineup, sitting above the existing $20 Plus tier. The headline perk is five times more Codex usage than Plus, aimed at developers running longer, more intensive coding sessions.

Pro subscribers still get everything the current Pro plan offers, the exclusive Pro model, plus unlimited access to Instant and Thinking models.

To sweeten the launch, OpenAI is temporarily doubling down on that promise: through May 31st, new Pro subscribers get up to ten times the Codex usage of Plus. A limited-time nudge to get people building.

BEWARE: THESE METHOD CAN REDUCE MODEL PERFORMANCE

If you've spent any time running AI coding agents, you've hit the wall. The model stalls, the session cuts out, or your bill quietly doubles. Here are three ways to stop letting usage limits slow you down.

1. Switch to a smarter model pairing.

Running your most powerful model end-to-end on every task is overkill and it burns through your quota fast.

The advisor strategy, recently introduced by Anthropic, is a good example of a better approach. Let a cheaper model like Sonnet or Haiku handle the bulk of the work and only escalate to Opus when the task genuinely demands it.

You get near-flagship intelligence without the flagship price tag or the usage hit.

2. Let a tool manage your sessions for you.

Tools like Clauditor (not sponsored) watches your session in real time and rotates it out before it gets too expensive.

If a session starts burning through budget too fast, it blocks it, saves your full context, and injects it cleanly into a fresh session, so you pick up exactly where you left off without losing a single thread. No manual babysitting, no scrambling to reconstruct what the agent was doing mid-task.

Cache what your agent already knows.

A huge chunk of AI usage gets wasted re-reading the same context over and over. Your codebase structure, your conventions, your documentation.

Every new session starts cold and burns tokens just getting back up to speed. Prompt caching lets you store that static context so it doesn't need to be re-processed each time.

The agent hits the ground running, your sessions stay leaner, and you get significantly more useful work out of the same allowance.

We have a VERY exciting edition for our next newsletter, until then!