Start Here
The Model Fleet
The models currently serving camelStream, the intelligence floor they must clear, and the guarantees behind every stream
camelStream doesn't sell a model. It sells a floor. Your requests are served by a changing fleet of frontier models, and every model must clear the same intelligence floor before it can serve a stream. This page is the official, current list.
The intelligence floor
Every model in the fleet scores at or above 70% on Terminal-Bench 2.1 or 50 on the Artificial Analysis Intelligence Index. Below the bar, a model doesn't serve.
The current fleet
| Model | Family | Version | Verify the scores |
|---|---|---|---|
| DeepSeek V4 Flash | DeepSeek | Most recent publicly available version (0731 as of August 2026) | AA Intelligence Index · Terminal-Bench 2.1 |
| Gemini Flash | Gemini | Most recent publicly available version (3.7 as of August 2026) | AA Intelligence Index · Terminal-Bench 2.1 |
| GPT Luna | GPT | Most recent publicly available version (5.6 as of August 2026) | AA Intelligence Index · Terminal-Bench 2.1 |
| Muse Spark | Muse | Most recent publicly available version (1.2 as of August 2026) | AA Intelligence Index · Terminal-Bench 2.1 |
The fleet changes as new models launch and partnerships evolve. This page is updated when it does.
Last updated: August 24, 2026
What every stream is guaranteed
- The intelligence floor. Every fleet model scores at or above 70% on Terminal-Bench 2.1 or 50 on the AA Intelligence Index.
- Latest versions. When your stream is paired with a model, you get its most recent publicly available version, not a frozen snapshot.
- 260K context. Every request gets at least a 260K-token window, and more when the model serving you supports it.
- One generation per stream, always. Bursts of extra parallelism can happen when capacity allows, but only streams guarantee it.
Context beyond the guarantee
When a conversation grows past what the serving model supports, camelStream compacts the middle of the conversation first. Older messages are hollowed out while their envelopes stay in place, so roles and tool-call IDs keep lining up, and your original task and the latest turns stay intact.
Speed targets
| Metric | Target |
|---|---|
| Throughput | p10 at or above 40 tokens per second, p5 at or above 20 tokens per second |
| Time to first token | p95 under 5 seconds |
Throughput percentiles are floors: p10 at or above 40 tok/s means 90% of requests stream at 40 tok/s or faster. The first-token percentile is a ceiling: 95% of requests see a first token within 5 seconds. Targets are tuning goals, not a service-level agreement. See the camelStream Terms.
How the flat price works
camelStream sources capacity across the fleet and partner providers, wherever frontier-floor intelligence is most economical: bulk deals, new-model offers, and data-for-training arrangements. That sourcing is what subsidizes unlimited tokens at a flat price. Requests and responses may be retained and used to train AI models, by us or by our inference providers. Your account details never are. The camelStream Terms and Privacy Policy describe the data rights in full.