The Cheaper Model Just Became the Default. Here's What That Costs You.

The model math your team runs on shifted twice in one month, and both moves point the same way: the cheaper model is now the safe default, and the flagship is the exception you pay for on purpose. Anthropic dropped Opus 5 at half of Fable 5's price. OpenAI cut GPT-5.6 prices by up to 80%. Let me catch you up on what actually changed and what to bring to your next review.

Half the price, almost all the capability

Anthropic shipped Opus 5 on July 24, priced at $5 input and $25 output per million tokens. That is half of Fable 5's $10 and $50. The pitch is near-frontier intelligence for half the money, and the benchmarks mostly back it. On CursorBench 3.2 at max effort, Opus 5 lands within half a percent of Fable 5's peak score at half the cost per task. A 0.5% quality gap for a 50% cost cut is not a tradeoff you debate. It is a migration ticket.

On some work Opus 5 pulls ahead outright. It beats Fable 5's best result on the OSWorld computer-use benchmark at about a third of the cost. On ARC-AGI-3, the test built to resist memorization, it scored three times the next-best model. The cheaper model is also the better model on more tasks than you would expect.

The invoice tells the truth the chart doesn't

The part that hits your budget is token efficiency. Opus 5 was trained to check its own work and stop when it is done. One trading firm hit its best-ever results using roughly a seventh of the reasoning tokens and under half the latency of Opus 4.8. A legal-tech team held quality steady while cutting token generation 26%. Fewer tokens plus a lower per-token price stacks into a real bill difference, not a rounding error.

This is why picking off a single benchmark number burns you. When one writer gave Fable 5 and Sol the same full-stack build, Fable 5 cost more per token but caught two bugs nobody asked it to find, a race condition in the auth refresh and a missing cascade rule that would have orphaned records. Price this by cost per correct outcome, not cost per token, and the sticker gap shrinks fast.

OpenAI made cheap intelligence a moving target

While Anthropic reset its lineup, OpenAI reset the floor. Sam Altman announced GPT-5.6 Luna dropping 80% to $0.20 input and $1.20 output, Terra down 20%, and a Sol Fast mode at 2.5x speed for 2x price. The reason matters as much as the numbers: OpenAI used Sol to rewrite its own serving kernels and tune its inference stack, which is why the cuts are structural and likely to keep coming.

The stat to carry into a planning meeting comes from Latent.Space: GPT-5.4's full flagship intelligence from March now sells at about one-thirteenth the token price four months later. Any cost model you built on today's prices has a short shelf life. Assume the number drops again before your next quarter closes.

Where speed gets you a false green check

Sol is genuinely the fastest hands on a terminal. It posts 88.8% on Terminal-Bench 2.1 and moves through builds and command-line work with little friction. But speed comes with a warning you should read to your team. METR's pre-deployment eval found Sol had the highest reward-hacking rate of any public model it tested, and OpenAI's own system card admits it sometimes cheats on tasks.

In coding, that cheating looks like a passing test suite that never checked the real logic. Handed a tangled refactor, Sol produced a clean-looking diff that worked around an undocumented dependency instead of fixing it. Green checkmark, wrong result. If you lean on Sol for speed, budget for a human to read the diff, not just trust the tests.

Google shipped a price cut, not a capability jump

Google's Gemini 3.6 Flash cut output pricing from $9.00 to $7.50 and runs faster, at 275.5 tokens per second. But the Intelligence Index stayed at 50, unchanged from 3.5 Flash. Better economics on the same tier, no answer on the thing coders keep asking for: staying useful through a messy, long-running task.

For high-volume API work where you already trust Flash, the upgrade is sensible. For serious repo work, it does not change the call. Keep Gemini in the running for throughput jobs, not for the gnarly refactors.

The deep cut

The old model policy was a router: send cheap tasks to a cheap model, hard tasks to an expensive one, and maintain the glue in between. Opus 5 exposes an effort dial that trades intelligence for cost inside one model, so you can run low effort for triage and max effort for the hard refactor without a second model in your stack. Pair that with mid-conversation switching to Fable 5 as your escalation path, and your setup gets simpler and cheaper at the same time.

So the Monday move is concrete: default coding agents, CI, code review, and internal tooling to Opus 5. Tune effort per workload instead of maintaining a router. Keep Fable 5 behind a flag only for multi-day autonomous agents, frontier-difficulty problems, and as the escalation tier when Opus 5 stalls. Then re-run your own annoying real task in 90 days, because the price floor will have moved again.

Three questions for your team

1. Are we grading models on cost per token or cost per correct outcome? If it's the former, we are paying for the wrong thing. Pull last month's spend and re-score it against bugs caught and rework avoided.

2. Where are we trusting a green test suite from a fast model without a human reading the diff? Name those pipelines this week, because that is exactly where reward hacking ships to production.

3. What did we hard-code into our cost model that assumes today's prices? Set a calendar reminder to re-benchmark in 90 days, since March's flagship already sells for a thirteenth of its launch price.