Cursor Killed Flat Auto Pricing — SpaceX Owns the Router

Cursor killed its flat Auto rate on August 24, 2026. The blended $1.25 input / $6 output per million tokens — the number that made Pro feel predictable at $20/month — is gone. Ten days earlier, SpaceX closed its $60 billion acquisition of Anysphere. The two events landing in the same month is not a coincidence. Your billing is now a function of which model Cursor’s router selects, and Cursor controls the router.

What Changed on August 24

Cursor announced the pricing shift via email on August 19 — no blog post, no changelog entry, no public documentation update. The pricing docs still listed the flat Auto rate as of publication. That alone tells you something about the communication posture here.

The mechanics:

• Old Auto: $1.25 input / $6 output per million tokens, regardless of which model the router selected. Predictable, if opaque.
• New Auto: Each request bills at the routed model’s list price. Composer 2.5 runs $0.50 input / $2.50 output — cheaper than the old flat rate. Claude Sonnet 5.5 and Opus 5.5 run substantially higher. Fable 5 and similar third-party routes are in the same expensive tier.

The allowance picture is deliberately murky. Cursor stated that included allowances for its own models increased — but published no dollar figures for Pro or Pro+ subscribers. You find out what your pool is by logging into your billing dashboard. Ultra subscribers have it in writing: prepaid credit allocations of $500 for third-party (frontier) model usage and $3,000 for Cursor-model usage, visible in the billing dashboard. Pro and Pro+ users are flying blind until they look.

Enterprise was the only plan Cursor extended any grace period to. The legacy Auto rate held for Enterprise through September 7, 2026. After that, same rules apply.

Why This Matters More Than a Pricing Tweak

The flat rate was a subsidy. Cursor was eating the spread when it routed your request to a frontier model at $4 input / $20 output (Opus 5.5) but charged you $6 output. That spread was a customer acquisition cost, and SpaceX acquisition economics ended it.

The structural change is bigger than the per-token numbers. By routing everyday work to Composer 2.5 and billing at its cheaper rate, Cursor creates a strong economic incentive for you to stay on Cursor-native models. But Auto still doesn’t tell you which model it called or what that request cost. You see the bill at the end of the cycle — not the routing decision that drove it.

Community reaction has been sharp. One Hacker News user reported $350 in overage in a single week — roughly $1,400 annualized against the “$20-ish” mental model most Pro subscribers carried. That’s not an edge case if your workflow routes heavy Agent or Ask sessions to Claude Sonnet 5.5.

The SpaceXAI angle adds a second layer of risk. A jointly developed SpaceXAI–Cursor model is in training on xAI’s Colossus supercomputing cluster, with a launch window described as imminent as of late August. When that model ships inside Cursor, routing decisions carry an ownership logic: SpaceX benefits when traffic moves to its own model instead of Anthropic’s or OpenAI’s. That’s not a conspiracy — it’s a normal vertical integration incentive. But it means “which model Auto picks” is now a question with a financial answer that isn’t purely about your code quality.

The Take

We’re not telling you to leave Cursor. It remains the strongest AI code editor for solo developers and small teams — the UX, the Composer workflow, the context window handling are all still ahead of the field. That position hasn’t changed because SpaceX bought the company.

What has changed is the predictability contract. Flat Auto was a deal. Per-model Auto is a market rate with Cursor as intermediary. That’s a meaningful difference for anyone who doesn’t check their billing dashboard monthly.

Here’s what to do before your next cycle:

• Log into your billing dashboard now. Find your actual pool size for Pro or Pro+. Don’t estimate.
• Review your last month’s Auto usage. If most requests hit Composer 2.5, you likely pay less than before — the math works in your favor.
• If you run heavy Agent sessions with frontier models daily, calculate your real per-month spend at Claude Sonnet 5.5 or Opus 5.5 list rates. If it clears $60/month, Pro+ is worth it. If it clears $200/month, evaluate Ultra — the prepaid credit allocations ($500 frontier, $3,000 Cursor-model) are published and fixed.
• Pin to Composer 2.5 explicitly where your workflow allows. Don’t rely on Auto to make the economical choice for you when you need a frontier model — invoke it intentionally.

Heavy terminal-workflow developers who were already running Claude Sonnet 5.5 separately from Cursor should do the math on claude-code as a standalone complement. Usage is priced independently from Cursor, and if you’re burning frontier model budget on both, you’re double-paying for the same capability.

openai-codex is the equivalent benchmark for OpenAI-native workflows — freemium, priced separately from any IDE, and worth a comparison if your Auto sessions are routing to GPT-tier models.

The routing is no longer neutral. SpaceX owns the model being trained to replace Anthropic in your editor. Watch the defaults. Your money follows the router, and the router now has a parent company with its own model on the way.