Featured
Published evals
Scores on MMLU, GSM-8K and HumanEval printed per release, with the full eval harness open so you can reproduce every number yourself.
100M+ requests a day
The reasoning model your engineers will actually trust.
Frontier capability at a fifth of the frontier price — and deterministic on the same input twice, so your evals mean something. Zero data retention by default, on every paid request.
Powering production at
Northwind
Helio Labs
Brightline
Foundry
Atlas Health
Lumen Co
Quickstart
The whole integration
Same request shape as the API you are already calling, so the migration is an import and a base URL.
summarise.tsts
import { Client } from "@your-org/sdk";
const client = new Client({ apiKey: process.env.API_KEY });
const res = await client.messages.create({
model: "reason-1",
temperature: 0, // deterministic: same in, same out
max_tokens: 1024,
tools: [summariseTicket],
messages: [
{ role: "user", content: "Summarise ticket #4821 for the on-call." },
],
});
console.log(res.content[0].text);Streaming is the same call with `stream: true`. Batch runs at half price with a one-hour window.
01CapabilitiesSystem card with every release
Built for production, not demos
The six things engineering teams check before they commit a roadmap to a model.
Deterministic on retry
Same input at temperature 0 is bit-for-bit the same output. Regression tests that actually regress.
Native tool use
Parallel function calling, JSON mode, and structured outputs that validate against your schema.
Streaming + batch
Token-level streaming for interactive loads; a batch API at half price for everything that can wait an hour.
Fine-tuning
LoRA fine-tunes from 50 examples. Hosted inference, and the weights are yours to export.
128K context
Long-context that holds attention at depth — needle-in-haystack scores published per release.
02At inferenceLive status at /status
- 100M+M
- Requests served daily
- 5×
- Cheaper than the leading frontier API
- 99.95%
- Uptime, trailing 12 months
- 0
- Days of data retention by default
Releases
Model + API changelog
Every model release carries a system card and a benchmark delta. Deprecations get six months and a migration guide.
- New
reason-1.2 — long-context recall improved at 128K
Needle-in-haystack scores published per depth band alongside the release. Same price.
- Improved
Batch API window cut from 24h to 1h
Half price, same models, and a hard completion guarantee inside the window.
- New
Structured outputs validate against your JSON schema
Invalid generations are retried server-side rather than returned for you to parse and discard.
- Security
Zero retention extended to the batch endpoint
Batch requests now match the interactive default: nothing stored after the response is delivered.
- Deprecated
reason-0.9 retires 18 Dec 2026
Six months of notice, a one-line migration, and both models served side by side until the date.
Why teams switch
The comparison your platform team will run anyway — here in the open.
Them
Frontier incumbents
- Input price / M tokens
$5–15
- Deterministic replay
Best-effort
- Data retention default
Opt-out, if offered
- Eval reproducibility
Marketing benchmarks
- Fine-tune weight ownership
Locked to platform
Us
Us
- Input price / M tokens
$1.20
- Deterministic replay
Bit-for-bit at temp 0
- Data retention default
Zero, on every paid request
- Eval reproducibility
Full harness published
- Fine-tune weight ownership
Exportable, yours
03$5 free credit on signup
Pricing
Pay-as-you-go per token. No commitments below the enterprise tier.
What engineering teams say
4.8·214 reviews
04More at /docs
Questions engineering teams ask
How does it compare on benchmarks?+
Every release publishes its MMLU, GSM-8K and HumanEval scores next to the previous one, at a fifth of the per-token cost of the frontier APIs. The eval harness is public — rerun every number yourself rather than taking ours.
Will the same prompt give the same answer twice?+
At temperature 0, yes — bit-for-bit. That determinism is what makes real regression testing possible.
What about data retention?+
Zero retention by default on every paid request. Enterprise can scope further, including fully in-VPC deployments for regulated workloads.
Is there a free tier?+
$5 of credit on signup, no card required. After that it's pay-as-you-go — no commitments below Enterprise.
What happens at the rate limit?+
429s with a Retry-After header and a real queue on Scale tier — not silent throttling. Burst headroom is documented per tier.
Can we fine-tune?+
LoRA fine-tunes from as few as 50 examples, hosted on the same infra. The resulting weights are exportable — they're yours.
Ship the feature this sprint.
API key in 60 seconds, $5 of credit, and docs your engineers won't swear at.
No card required · zero retention by default
Get started
An API key in 60 seconds — or tell us about your workload and we'll size a deployment.
- hello@example.com
- +1 (555) 000-0000
- 123 Main St, Newtown, CT