Gemini 3.6 Flash Is Live on Starchild: More Agent Work for the Same Budget
TL;DR: Gemini 3.6 Flash was launched earlier this week and is available on Starchild. It combines a 1 million-token context window, multimodal input, 251-token-per-second output, and an Artificial Analysis Intelligence Index score of 50.1. At $1.50 per million input tokens and $7.50 per million output tokens, it sits between cheaper GPT-5.6 Luna and more capable but slower Grok 4.5.
Here is how Gemini 3.6 Flash performs, how much it costs to run, and how it compares with other models in a similar price range.
More efficient agent loops
Google reports that Gemini 3.6 Flash uses 17% fewer output tokens than Gemini 3.5 Flash on the Artificial Analysis Index. On DeepSWE, the reduction reaches up to 65%. It also takes fewer reasoning steps and tool calls to complete multi-step workflows.
The Artificial Analysis Intelligence Index combines nine evaluations across mathematics, science, coding, and reasoning into one comparative score. In the latest leaderboard snapshot, Gemini 3.6 Flash ranks 15th with 50.1, narrowly behind Gemini 3.5 Flash and ahead of Claude Sonnet 4.6, Gemini 3.1 Pro Preview, and MiniMax M3.
| Rank | Model | Score | Rank | Model | Score |
|---|---|---|---|---|---|
| 1 |  Claude Fable 5 |
59.9 | 11 |  GPT-5.6 Luna |
51.2 |
| 2 | GPT-5.6 Sol |
58.9 | 12 |  GLM-5.2 |
51.1 |
| 3 |  Kimi K3 |
57.1 | 13 |  Muse Spark 1.1 |
50.6 |
| 4 | Claude Opus 4.8 |
55.7 | 14 |  Gemini 3.5 Flash |
50.2 |
| 5 | GPT-5.6 Terra |
55.0 | 15 | Gemini 3.6 Flash |
50.1 |
| 6 | GPT-5.5 |
54.8 | 16 | Claude Sonnet 4.6 |
47.2 |
| 7 |  Grok 4.5 |
53.8 | 17 | Gemini 3.1 Pro Preview |
46.5 |
| 8 | Claude Opus 4.7 |
53.5 | 18 |  Qwen3.7 Max |
46.0 |
| 9 | Claude Sonnet 5 |
53.4 | 19 |  MiniMax M3 |
44.4 |
| 10 | GPT-5.4 |
51.4 | 20 |  DeepSeek V4 Pro |
44.3 |
The model costs $1.50 per million input tokens and $7.50 per million output tokens. Its predecessor charged the same input price and $9.00 per million output tokens. The lower rate and reduced token use work together, which matters more than either number alone for agents that run repeatedly.
How it compares at a similar price
Gemini 3.6 Flash has real competition in this price band. GPT-5.6 Luna is cheaper and more concise. Grok 4.5 scores higher on intelligence but runs much slower. Claude 4.5 Haiku is less expensive, although it is a non-reasoning model with a smaller context window.
| Model | Input / output per 1M | AA Intelligence Index | Output speed | Context |
|---|---|---|---|---|
| Claude 4.5 Haiku |
$1.00 / $5.00 | 24 | 92 tok/s | 200K |
| GPT-5.6 Luna, max |
$1.00 / $6.00 | 51.2 | 162 tok/s | 1M |
| Gemini 3.6 Flash, high |
$1.50 / $7.50 | 50.1 | 251 tok/s | 1M |
| Grok 4.5, high |
$2.00 / $6.00 | 53.8 | 67 tok/s | 500K |
Each model has a clear operating profile:
- GPT-5.6 Luna: Choose it when minimizing token price and output volume matters most.
- Grok 4.5: Choose it for a higher measured intelligence score and lower output-token price. Gemini 3.6 Flash produces tokens almost four times faster and provides twice the context.
- Claude 4.5 Haiku: Choose it for simpler, latency-sensitive tasks that do not need deep reasoning or million-token context.
- Gemini 3.6 Flash: Choose it when an agent needs speed, long context, multimodal input, and substantial reasoning at the same time.
Our pick for most agent workloads is Gemini 3.6 Flash. GPT-5.6 Luna wins on price and Grok 4.5 scores higher on measured intelligence, but Gemini offers the strongest overall balance of speed, context, multimodality, reasoning, and cost. For a general-purpose agent that moves between research, coding, documents, and tools, that balance matters more than winning one metric.
Where it fits on Starchild
Gemini 3.6 Flash accepts text, images, speech, and video. Google lists native support for thinking, function calling, and Computer Use, with a maximum output of 64,000 tokens. That makes it useful for:
- Coding agents that inspect, modify, and test repositories
- Research agents working across documents, screenshots, and recordings
- Browser and interface automation using Computer Use
- Scheduled monitoring where lower cost compounds across repeated runs
- Multi-agent systems that need fast sub-agent loops
Starchild supplies the environment around the model: persistent memory, connected apps, APIs, files, code execution, scheduled tasks, and reusable skills. You can use Gemini 3.6 Flash as the workhorse for frequent or output-heavy steps, then switch to another model when a task needs a different balance of price, speed, or reasoning.
Gemini 3.6 Flash is available on Starchild now. Select it in your agent's model settings, or use Conductor Mode to let Starchild choose based on the task.
Try Gemini 3.6 Flash on Starchild
Sources: Google's Gemini 3.6 Flash announcement, Google's current model documentation, and Artificial Analysis pages for Gemini 3.6 Flash, GPT-5.6 Luna, Grok 4.5, and Claude 4.5 Haiku.