The Signal · Issue 01 · Frontier Models
The falling price of a capable token
Anthropic's Haiku 5.5 and xAI's Grok 4.7 both lead with price and speed. Cheaper capable models change what is worth automating, and what is worth measuring.
Facts are attributed and linked to their sources. The analysis and opinions are the author's own.
What it is
On 7 October Anthropic introduced Claude Haiku 5.5, which it calls the cheapest, fastest and most capable small model it has released. Alongside it, Anthropic halved the price of Claude Sonnet 5.5 cache reads, from $0.20 to $0.10 per million tokens, which it says makes Sonnet 5.5 around 20% cheaper on most agentic work, and added computer-use and browser-use support in beta to its Python and TypeScript SDKs.
On 21 September, according to its API release notes, xAI made Grok 4.7 available on its API, with a 500k context window, priced at $2 per million input tokens and $6 per million output tokens below 200k prompt tokens. xAI says Grok 4.7 uses a larger base model than Grok 4.6, was trained with a longer reinforcement-learning run weighted toward many-hour tasks, and is better at verifying its own work.
Why it matters
When capable inference gets cheaper, the economics of automation shift. Tasks that were not worth a model call become worth several, and the natural next step is many cheap sub-agents checking each other's work. That is efficient, and it multiplies the number of decisions nobody looked at.
Cheaper tokens multiply the number of decisions nobody looked at. Budget for the looking.
There is also a measurement trap. Launch posts lead with benchmark tables, and every lab picks the ones that flatter it. Goodhart's warning applies: once a benchmark becomes the target of a release, it says less about the thing we care about. The honest test is a small set of your own tasks, with known answers, run on every model you consider.
The practical consequence is architectural. When small models get cheap and good, the pattern shifts from one large model doing everything to a larger model planning and several small ones executing. Anthropic positions Haiku 5.5 for exactly this kind of sub-agent and high-volume work, and adding computer use and browser use to its SDKs points the same way.
That architecture is efficient and harder to audit. Each sub-agent call is a decision, and the number of decisions grows quickly. If you cannot sample and review them, you cannot say whether the system is doing what you think. I would budget a fixed share of any AI workload's cost for evaluation and review, the way we budget for model monitoring in credit risk.
On benchmarks, the honest position is modest. Vendor tables are useful for shortlisting, not for deciding. A private evaluation set of fifty to a hundred of your own tasks, with known good answers, refreshed every quarter, will tell you more about fit than any public leaderboard.
What I'd watch
- Total cost per completed task, not per token. Self-verification and retries can change the bill more than list prices.
- Whether teams keep a private evaluation set and refresh it, so vendor benchmarks are context rather than the decision.
- How small, fast models are governed when they run as sub-agents inside a larger workflow.
Sources
- Introducing Claude Haiku 5.5. Anthropic, 7 Oct 2026.
- Introducing Grok 4.7. xAI (x.ai), 21 Sep 2026.
- Release notes. xAI API docs, 21 Sep 2026 entry.