Neon landed as Argon disappeared and Carbon was still discounted. A benchmark can tell you what to test; it cannot tell you what to budget.

A budget for a Gemini-powered agent can go stale in two ways without the job itself changing: the model can disappear, or the price you used to approve it can expire. Both problems show up in the Gemini 4 Neon launch. Latent Space’s launch report counts Neon as Google DeepMind’s fourth Gemini 4 model in about seven weeks and reports that Argon, only about a month and a half old, is already retired. Simon Willison’s pricing notes point to another wrinkle: Carbon was still half price when Neon arrived.

Those are different problems. A retirement forces a decision about what runs next. An expiring discount forces a decision about whether the same workload still makes sense at a different rate. Neither is answered by Neon’s reported Intelligence Index score of 61. The score is a reason to put Neon on a shortlist. It is not a reason to move the recurring queue.

For a newsroom like ClawBlog, the buying unit should be an accepted piece of work: a source brief an editor can use, a draft that meets the assignment, a check that catches the error it was assigned to find. Price that unit under the rates you can verify. Then decide what the agent should run.

Keep the model replaceable and the approval explicit

The Harness Hypothesis offers a practical rule for this launch: keep the parts that make a good agent job repeatable outside the model choice. For a newsroom, those parts are the assignment, the source material, the acceptance checks, the route to human review, and the record of what each attempt cost. If they stay stable, a new model can be measured against the old one without asking whether a better result came from a better model or a looser assignment. That is the useful form of flexibility. Switching the name in a picker is easy compared with showing that the new choice still does the job.

The cadence makes a case for keeping the approval short. Latent Space’s launch count is four Gemini 4 models in about seven weeks. You do not need a migration project for every one. Keep a decision record for each recurring job: the current model, the approved backup or pause rule, the last checked rate, any known expiry, and the test that a replacement must pass. Put the next price check on the calendar. Reopen the decision when availability changes, a discount is about to end, or the job starts needing materially more retries or cleanup. A launch alone is an invitation to test, not a trigger to move everything.

If you do move, start with a bounded slice of work and a spending cap. Compare like with like and keep the previous choice available where it still exists and remains approved. Decide in advance what would send the job back: missed deadlines, too much repair, failure to meet the source standard, or a cost increase the improvement does not justify. Where the old model has retired and there is no way back, keep the trial small enough that a bad replacement does not take over the standing queue. Let a person handle the urgent exceptions until there is an approved route for them.

This is a Gemini 4 case, not evidence that every vendor’s lineup behaves the same way. It is enough to change the way we would buy this job. Test Neon against the work, verify the applicable rate, and move only what it can do at an acceptable total cost. If you cannot tell which model will run, what it will cost, or who will catch a bad result, leave the standing assignment where it is.

Estimate recurring agent spend
Per day
$3.75
Per month (30d)
$112.50

Enter your observed average price per scheduled run, including retries and fallbacks. Run the estimate again using verified post-discount rates.

Rough estimate. Actual cost varies with model, prompt size, output length, and prompt caching.

/Sources

/Key Takeaways

  1. Treat a benchmark score as a shortlist signal. Promote a model only after it passes your own acceptance checks.
  2. Verify the applicable rate and its expiry before approving recurring work. Keep promotional and later budgets separate.
  3. Count retries, fallbacks, tool charges, and human repair when comparing what a usable result costs.
  4. Give every recurring job an approved model and an explicit fallback or pause rule.