The frontier model field guide: what you actually pay per million tokens
Context windows, list prices, latency under load and the lock-in clauses nobody reads. I ran the same twelve-task suite across nine flagship models.
Issue No. 14 · The Business AI Beat
The AI stack got boring. That is the best news operators have had all year.
Eighteen months of demos have finally given way to procurement, invoices and postmortems. I spent the last quarter inside eleven deployments — here is what actually survived contact with a P&L.
Two Novembers ago I sat in a boardroom while a vendor promised that an agent would "own" a client’s accounts-payable function by spring. It did not. What it did do — and this is the part nobody puts on a slide — was reduce a nine-step invoice exception workflow to three, and cut the median handling time from eleven minutes to four. The finance director called that a failure for a quarter, then quietly renewed.
That gap between the pitch and the payoff is the whole beat now. The frontier labs ship a new flagship roughly every eleven weeks; the interesting question stopped being which model wins some time in 2025. It is now: what does inference cost you at volume, how much of your engineering week is spent babysitting retries, and what happens to your pricing when your vendor raises output-token rates by forty per cent with thirty days’ notice?
I test with my own money and my own logs. Every claim in this issue comes with a number attached, or it does not run. Where I have not tested something myself, I say so. And where a vendor refused to give me a straight answer on per-seat versus per-outcome pricing — which happened three times this quarter — I say that too.
In this issue — five dispatches
Agents that ship, and the six that quietly died in staging
LangGraph, the OpenAI Agents SDK, CrewAI, Google ADK and Pydantic AI, judged on the only metric that matters: what happens on step fourteen.
Services-as-software: when you buy the outcome, not the seat
Per-resolution support, per-matter legal review, per-lead outbound. I priced four outcome-billed vendors against the headcount they claim to replace.
Zapier, Make, n8n, Workato: the task-pricing trap
Five platforms, one identical eleven-step workflow, and a monthly bill that ranged from £0 to £1,940. The spread is not about features.
The unglamorous layer: gateways, evals, sandboxes and traces
Nobody demos an LLM proxy. But routing, observability and eval tooling are where the last three deployments I reviewed were won or lost.
Standfirst — How this desk works
I do not run vendor benchmarks, and I do not take briefing-only stories
Every piece on this site starts from a live workload. For model coverage that means a fixed suite of twelve tasks — four extraction, three long-context summarisation, two code-repair, two tool-calling chains and one adversarial refusal test — run at temperature 0.2, five times each, from a London region, on a weekday morning when the shared endpoints are busiest. Cherry-picked 3 a.m. latency numbers are marketing, not measurement.
For platform coverage I build the same eleven-step workflow every time: webhook in, enrichment lookup, conditional branch, LLM classification, CRM write, Slack notification, error handler, retry, dead-letter queue, daily digest, audit log. It is deliberately mundane. Mundane is what companies run.
I publish list prices with the date I checked them, because they move. I flag when a vendor has given me trial credits — it has happened twice and both times I said so in the piece. And when I get something wrong, the correction sits at the top of the article with the original text struck through, not quietly patched at 2 a.m.
On independence
viblyhappybundle.homes takes no vendor sponsorship, no affiliate commission and no paid placement. The publication is funded by reader subscriptions and a small consulting practice that never advises on tools covered in the same quarter. If that changes, it will be disclosed here first, in this section, in this typeface.