Chapter §01 · The State of AI Inference Reputation, 2026
The Landscape
Compute got cheap, models got good, and the last scarce commodity in machine intelligence turned out to be an old one: trust. A field report on who has earned it, who is buying it with uptime, and who is still writing cheques against a reputation they do not yet possess.
- At a glance
- The field's mean composite reputation stands at 83 / 100 across 20 providers.
- OpenAI leads the reputation index; Cerebras Inference leads on raw throughput (~520 tok/s).
- 18 of 20 venues offer open-weight model access — the centre of gravity has shifted toward open models.
- On indicative output price, cheapestinference is the lowest in our set at $0.35 / 1M tokens.
Begin with the thing everyone can measure, because it is the thing that no longer decides anything. Price per million tokens has fallen so far, so fast, that across the open-weight venues it now clusters within a few tens of cents of itself. A buyer who agonises over a two-cent spread on output tokens is optimising the wrong variable. The spread that matters — the one that separates a good year from a bad outage post-mortem — is the spread in reputation.1
That is the argument of this report, and the reason Datum Index exists in the form it does. We could have built another latency leaderboard. The world has enough of those, and they measure a moment rather than a relationship. What the world lacks is a considered, editorial reading of which providers have earned the right to carry production traffic — and an honest account of how confident anyone can be in the numbers those providers publish about themselves.
The shape of the field in 2026 is a barbell. At one end sit the frontier labs and hyperscalers, trading on model quality, enterprise diligence and years of accumulated uptime; they command the highest composites in our index and defend them with compliance attestations that smaller rivals cannot yet match. At the other end sits a vigorous population of aggregators, routers and silicon specialists competing on economics and speed. The middle — where a provider is neither the safest nor the cheapest — is the hardest place to stand.
The silicon specialists deserve a paragraph of their own, because they have rewritten one column of the table. Custom inference hardware has pushed reference throughput past the 520-token-per-second mark and time-to-first-token below 110 milliseconds — figures that would have read as typos two years ago.2 Speed, though, is not the same as trust: several of the fastest venues in our set carry thinner compliance coverage and shorter public track records, and their reputations reflect the discount.
A final structural note. 18 of the 20 venues we cover now offer open-weight model access, most of them behind an OpenAI-compatible API. The practical consequence is that switching cost has collapsed for a large slice of the market. When any competent aggregator can serve the same open model behind the same API dialect, the provider's reputation — its record of being there, billing honestly, and handling data as promised — becomes almost the entire basis for choosing one over another.
When switching cost collapses to a base-URL change, reputation stops being a soft factor and becomes the product. The State of Inference 2026
The Commented Benchmark
Six providers, read against each other
Not a ranking of speed, but a reading of it. Each row pairs the measurable figures with our editorial verdict — and a mark for how far to trust the numbers.
| Provider | Category | Out $/M | Tok/s | TTFT | Score | Verdict | Data |
|---|---|---|---|---|---|---|---|
| OpenAI | Frontier Lab | $10.00 | 90 | 480ms | 91 | Benchmark of the field | high |
| Anthropic | Frontier Lab | $15.00 | 78 | 520ms | 90 | Benchmark of the field | high |
| OpenRouter | Router | $0.90 | 120 | 400ms | 88 | Highly commended | high |
| Cerebras Inference | Specialist Hardware | $0.80 | 520 | 110ms | 82 | Recommended with confidence | medium |
| cheapestinference | Subscription (flat-rate) | $0.35 | 120 | 340ms | 83 | Recommended with confidence | seed |
Read down the throughput column and the specialists win it outright; read down the score column and the frontier houses reclaim the lead. That inversion is the whole story of 2026 in two columns. Raw tokens per second is a solved problem for anyone willing to buy the right silicon; a reputation for being there on a bad day is not for sale at any clock speed.
Note, too, the confidence marks. Several of the most striking figures in the table carry a seed flag — they are illustrative of the shape of an offering, not verified quotes. We would rather show the shape and mark our uncertainty than launder an estimate into a fact.3
The Verdict
Where the ground is shifting
Our expectation for the coming year is not a price war — that war is essentially over, won by physics and open weights — but a trust war. The providers who pull ahead will be the ones who publish honest status histories, earn the unglamorous compliance attestations, and price transparently enough that a buyer never feels surprised by an invoice. The composite scores in our index will move in response to exactly those behaviours, because that is what our readers, and the practitioners behind our reviewer notes, actually reward.
Datum Index will keep the record. We will revise these figures as first-party verification allows, upgrade or downgrade our confidence marks in public, and never pretend to a precision we do not have. That, in the end, is the only reputation an observatory can afford to protect: its own.
- Indicative prices in this report are approximate USD per 1M output tokens for a representative open model where the provider hosts one; proprietary-only providers are listed by flagship model. They are not quotes. Blended prices are approximate USD per 1M tokens for a representative open model where the provider hosts one; proprietary-only providers list a flagship model instead. Latency figures are order-of-magnitude reference points, not SLAs.
- Latency and throughput figures are order-of-magnitude reference points drawn from public sources, not service-level guarantees, and will vary by model, region, context length and load.
- reputation.score is an editorial composite index (0-100), not a first-party metric. See each site's Methodology page for the weighting model. The confidence legend is reproduced in full on our Methodology page.