Production models for agent workloads
Built for the reasoning a loop demands, and priced for running it all day. Same SDK, one line of configuration.
Start with $25 in free credit. No card required.


Our top‑tier models optimized for production workloads
Hal reasons like a frontier model, so the loop does not get longer.
Hal 1.0
Comparable to Anthropic
Opus 5.5 or OpenAI GPT-6 Astra
Clarke 1.0
Comparable to Anthropic
Sonnet 5.5 or OpenAI GPT-6 Sol
Tycho 1.0
Comparable to Anthropic Haiku
5.5 or OpenAI GPT-6 Luna
Every tier, priced next to the models it stands in for.
Radium
Anthropic
OpenAI
Tycho 1.0
$0.50 / $2.25
4.7x Haiku 5.5
Claude Haiku 5.5
$0.10 / $0.50
GPT-6 Luna
$0.10 / $0.50
Clarke 1.0
$1.50 / $7.00
28% less
Claude Sonnet 5.5
$2.00 / $10.00
GPT-6 Sol
$2.00 / $10.00
Hal 1.0
$2.25 / $13.75
35% to 74% less
Claude Opus 5.5
$4.00 / $20.00
GPT-6 Astra
$10.00 / $50.00
Input / output, USD per million tokens. Anthropic list prices as published on 12 October 2026; Haiku 5.5 shown for prompts up to 100K tokens. OpenAI GPT-6 standard rates as announced in September 2026, for prompts under 272K tokens. Savings use a 70% input / 30% output blend.
Scored against frontier models on real product tasks
Independent Radium benchmark, not endorsed by Anthropic. Claude and Opus are Anthropic PBC trademarks, used for identification only. Results are a single run, provided “as is,” and not a guarantee of future performance. See the full benchmark report
Control the cost per completed task
Every new account gets $25 in credit. No card required.

What changes
The base URL in your client
The model string
The invoice
What stays the same
The agent loop
The Anthropic or OpenAI SDK
Tool and function calls
Parallel tool calls
Streaming, including tool deltas
Structured output
Retry and error handling
Evals, tracing and observability
What the difference looks like monthly
Same tasks, different rate, two bills to compare.
Hal 1.0 against Claude Opus 5.5 list pricing, using real per-task token averages from the production benchmark. See the working
Six cost layers sit inside a frontier token. Two sit inside ours
The same arithmetic holds whether the call comes from an agent loop or a retrieval chain.


The switch is one line, and so is the way back

That is the right question, and it is why we publish step counts next to prices. In the measured run Hal took thirty-three steps against thirty-one and finished the same task, so loop length held and the rate difference carried through. Run the same comparison on a task of your own before you believe ours.
The same thing that happens today. Your loop catches it, retries and carries on, and the error codes your retry logic keys on are unchanged.
Context window and maximum output are published in the model reference. Transcripts that need compaction get compacted the way they do now, because the request shape has not changed.
Time to first token is comparable, and sustained throughput is what matters once a loop is running. Both are in the benchmark results, measured across a full task rather than a single call.
If it takes a base URL, yes. LangGraph, LlamaIndex, CrewAI, Pydantic AI, the OpenAI Agents SDK, Vercel AI SDK and LiteLLM all work by changing provider config, and there is a written guide for each.
No. Tool calling, parallel tool calls and streaming tool deltas behave the same way, because the API is compatible rather than similar.
No, and agent system prompts are usually the most tuned thing a team owns. Re-run your evals as a matter of course rather than because we expect a difference.
Published, and above what most agent traffic needs. Agents burst rather than trickle, so if your ceiling is the problem we raise it rather than throttle you, because we own the capacity.
Point a staging agent at the endpoint and run the tasks you already run. The harness does it for you and prints cost per completed task on both endpoints.
Remove one environment variable. Nothing in the loop changed on the way in.


