Raven Flash vs GPT-5.6 Terra, Gemini 3.7 Flash and Claude Opus 4.8
In our published results, Raven Flash scores higher than Gemini 3.7 Flash on all eight tests and higher than GPT-5.6 Terra on seven of eight. Against Claude Opus 4.8, it is ahead on three, level on the Intelligence Index and behind on four. It costs $0.50 per 1M input tokens through the DipoleML API.
Key facts
- Ahead of Gemini 3.7 Flash on all eight tests of our published results.
- Ahead of GPT-5.6 Terra on 7 of 8; GPT-5.6 Terra is ahead on DeepSWE v1.1.
- Against Claude Opus 4.8: ahead on Terminal Bench 4.0, AutomationBench and DeepSWE v1.1, level on the Intelligence Index, behind on the other four.
- First among the models compared on Terminal Bench 4.0 (34.1) and AutomationBench (49.2), and level first on the Intelligence Index.
- $0.50 per 1M input tokens, $0.05 cached and $2.50 output; results company-reported, evaluated with maximum reasoning effort.
Raven Flash is our high-velocity model, built for low latency, responsive autonomous agents and high-frequency code execution. In our published results it is measured on eight tests against a table of other models. This piece takes three of them and sets each one beside Raven Flash in turn. The results are company-reported, from our technical report of October 2026, and every test was evaluated with maximum reasoning effort.
Which models are compared here?
- GPT-5.6 Terra is one of the three models in OpenAI's GPT-5.6 family, with Sol and Luna. OpenAI's API documentation describes it as a GPT-5.6 model that balances intelligence and cost.
- Gemini 3.7 Flash is a model in Google's Flash series; Google introduced it on August 13, 2026. Google DeepMind's model card places it in the Gemini 3 model family.
- Claude Opus 4.8 is a model from Anthropic, which released it on May 28, 2026, as a new version of Claude Opus.
All three appear beside Raven Flash on every one of the eight tests, which is why we set them side by side here.
How do the four compare on each test?
Here are the eight results in one table, with the highest of the four on each test in bold.
| Benchmark | Raven Flash | Claude Opus 4.8 | GPT-5.6 Terra | Gemini 3.7 Flash |
|---|---|---|---|---|
| Terminal Bench 4.0 | 34.1 | 21.7 | 24.6 | 17.5 |
| Agent's Last Exam | 26.6 | 27.0 | 25.1 | 22.4 |
| Automation | 49.2 | 41.0 | 37.2 | 38.6 |
| GDPVal-AA | 1,651 | 1,731 | 1,571 | 1,527 |
| Deep | 65.4 | 58.0 | 69.6 | 65.3 |
| Intelligence Index | 43 | 43 | 37 | 41 |
| HLE w/ Tools | 55.8 | 57.9 | 52.1 | 49.8 |
| AA-Briefcase v1.1 | 1,462 | 1,480 | 1,385 | 1,132 |
The chart draws the same results as bars, one test at a time.
Raven Flash and the three models, test by testScore
| Model | Score |
|---|---|
| Raven Flash | 34.1 |
| GPT-5.6 Terra | 24.6 |
| Claude Opus 4.8 | 21.7 |
| Gemini 3.7 Flash | 17.5 |
| Model | Score |
|---|---|
| Claude Opus 4.8 | 27.0 |
| Raven Flash | 26.6 |
| GPT-5.6 Terra | 25.1 |
| Gemini 3.7 Flash | 22.4 |
| Model | Score |
|---|---|
| Raven Flash | 49.2 |
| Claude Opus 4.8 | 41.0 |
| Gemini 3.7 Flash | 38.6 |
| GPT-5.6 Terra | 37.2 |
| Model | Score |
|---|---|
| Claude Opus 4.8 | 1731 |
| Raven Flash | 1651 |
| GPT-5.6 Terra | 1571 |
| Gemini 3.7 Flash | 1527 |
| Model | Score |
|---|---|
| GPT-5.6 Terra | 69.6 |
| Raven Flash | 65.4 |
| Gemini 3.7 Flash | 65.3 |
| Claude Opus 4.8 | 58.0 |
| Model | Score |
|---|---|
| Raven Flash | 43 |
| Claude Opus 4.8 | 43 |
| Gemini 3.7 Flash | 41 |
| GPT-5.6 Terra | 37 |
| Model | Score |
|---|---|
| Claude Opus 4.8 | 57.9 |
| Raven Flash | 55.8 |
| GPT-5.6 Terra | 52.1 |
| Gemini 3.7 Flash | 49.8 |
| Model | Score |
|---|---|
| Claude Opus 4.8 | 1480 |
| Raven Flash | 1462 |
| GPT-5.6 Terra | 1385 |
| Gemini 3.7 Flash | 1132 |
Evaluated with maximum reasoning effort enabled. Source: Dipole ML Inc. Technical Report (October 2026).
Where is Raven Flash first?
Among all the models in its table, Raven Flash is first on Terminal Bench 4.0 (34.1) and AutomationBench (49.2). It is level first on the Intelligence Index at 43, with Claude Opus 4.8 and Muse Glimmer. On the other five tests it is second.
How does Raven Flash compare with Gemini 3.7 Flash?
Raven Flash scores higher than Gemini 3.7 Flash on all eight tests. The narrowest gap is on DeepSWE v1.1, 65.4 to 65.3. Relative to the score, the widest is on Terminal Bench 4.0, 34.1 to 17.5.
How does Raven Flash compare with GPT-5.6 Terra?
Raven Flash is ahead of GPT-5.6 Terra on seven of the eight: Terminal Bench 4.0 (34.1 to 24.6), Agent's Last Exam (26.6 to 25.1), AutomationBench (49.2 to 37.2), GDPVal-AA (1,651 to 1,571), the Intelligence Index (43 to 37), HLE w/ Tools (55.8 to 52.1) and AA-Briefcase v1.1 (1,462 to 1,385).
GPT-5.6 Terra is ahead on DeepSWE v1.1 (69.6 to 65.4), a gap of 4.2. It is first among the models compared on that test, and Raven Flash is second.
How does Raven Flash compare with Claude Opus 4.8?
This is the closest of the three. Raven Flash is ahead on Terminal Bench 4.0 (34.1 to 21.7), AutomationBench (49.2 to 41.0) and DeepSWE v1.1 (65.4 to 58.0). The two are level on the Intelligence Index at 43.
Claude Opus 4.8 is ahead on the other four: Agent's Last Exam (27.0 to 26.6), GDPVal-AA (1,731 to 1,651), HLE w/ Tools (57.9 to 55.8) and AA-Briefcase v1.1 (1,480 to 1,462). It is first among the models compared on each of them, with Raven Flash second. Of all the models in Raven Flash's table, it is the only one ahead of Raven Flash on more tests than Raven Flash is ahead of it. Its narrowest lead, relative to the score, is on AA-Briefcase v1.1.
A bar chart of five scores on AA-Briefcase v1.1, with Claude Opus 4.8 first at 1,480 and Raven Flash second at 1,462.
AA-Briefcase v1.1Score
| Model | Score |
|---|---|
| Claude Opus 4.8 | 1480 |
| Raven Flash | 1462 |
| GPT-5.6 Terra | 1385 |
| DeepSeek-V4-Vision | 1266 |
| Gemini 3.7 Flash | 1132 |
Evaluated with maximum reasoning effort enabled. Source: Dipole ML Inc. Technical Report (October 2026).
Which other models are in Raven Flash's table?
Each test sets Raven Flash beside four other models. Besides the three above, two more models appear in the table, each on fewer tests:
- DeepSeek-V4-Vision appears beside Raven Flash on seven tests, and Raven Flash scores higher on all seven.
- Muse Glimmer appears beside Raven Flash on one test, the Intelligence Index, where the two are level at 43.
What does Raven Flash cost?
Through the DipoleML API, Raven Flash bills $0.50 per 1M input tokens, $0.05 per 1M cached input tokens and $2.50 per 1M output tokens. It reads up to 1M tokens of context and writes up to 128K tokens in one answer. It also runs in the Raven app and in DCode on your plan; every plan and price is on the pricing page.
How can I try Raven Flash on my own work?
Benchmarks tell you where to start, and your own tasks tell you the rest. Run Raven Flash in the playground, or send it requests through the DipoleML API with the model name Raven Flash. If the work runs over many dependent steps, our guide to choosing between Raven Max and Raven Flash says when to move it to Raven Max, whose own head-to-head is in Raven Max vs GPT-5.6 Sol and Fable 5. Every result, test by test, is on the Raven Flash page.
Product names are trademarks of their owners and are used here only to identify the products.
Questions and answers
Who ran these tests?
We did. These are our published results, company-reported, from our technical report of October 2026, and every test was evaluated with maximum reasoning effort. The same results are on the Raven Flash page.
Is Raven Flash ahead of Claude Opus 4.8?
On three of the eight tests: Terminal Bench 4.0, AutomationBench and DeepSWE v1.1. The two are level on the Intelligence Index, and Claude Opus 4.8 is ahead on the other four, with Raven Flash second on each of them.
Why are these three models compared with Raven Flash?
They are in Raven Flash's table in our published results, and all three appear beside it on all eight tests. Each test sets Raven Flash beside four other models.
What does Raven Flash cost?
Through the DipoleML API, $0.50 per 1M input tokens, $0.05 per 1M cached input tokens and $2.50 per 1M output tokens. It also runs in the Raven app and in DCode on your plan.
First published Last updated
