Back to changelog

Gemini 3.8 Flash, Muse Spark 1.3 & More Models

Gemini 3.8 Flash lands on Google AI Studio and Vertex AI with a 1M context at $0.75/M input, Meta's Muse Spark 1.3 arrives with a 1M context and a $0.10/M Contributor tier, and Kimi K3, GLM-5.3 Flash, and Qwen3.8 Flash pick up new deployments across Runware, Novita, SCX.ai, and Alibaba Cloud.

A cluster of glowing new model chips shaped as a lightning bolt, a sparkle, and a picture frame on a circuit board, surrounded by brains, lenses, and rockets

Two weeks of catalogue work in one place: a new Google Flash generation on both Google clouds, Meta's next Muse Spark, and extra deployments for the open-weight models people route the most.

Gemini 3.8 Flash

1google-ai-studio/gemini-3.8-flash2google-vertex/gemini-3.8-flash
Spec Value
Context 1,048,576 tokens
Max output 65,536 tokens
Input $0.75 per 1M tokens
Cached input $0.075 per 1M tokens, 4,096-token cache minimum
Output $3.75 per 1M tokens
Reasoning low, medium, high
Inputs Text, image, audio, document
Also Tools, JSON schema output, web search, Flex and Priority tiers

Both mappings were verified live, including cache reads and hand-checked costs. Completion tokens include reasoning tokens.

Muse Spark 1.3

muse-spark-1.3 brings Meta's agentic reasoning model to the gateway in beta: a 1M-token context, reasoning effort from minimal to xhigh, vision, tools, and structured JSON, at $1.25/M input, $0.15/M cached input, and $4.25/M output.

muse-spark-1.3-contributor is the same model at $0.10/M input, $0.002/M cached input, and $0.20/M output for workloads that allow Meta to train on their data. It is modelled as a separate, training-eligible provider, so a compliance policy that blocks API training never routes to it unless you explicitly allow it.

More Deployments for the Open-Weight Favorites

  • Kimi K3 on Runware (runware/kimi-k3): 1M context, reasoning none through max, vision, tools, JSON schema, and prompt caching at $3/M input, $0.30/M cached, $15/M output.
  • GLM-5.3 Flash (glm-5.3-flash): Z.ai's lightweight 1M-context coding model with vision and low, high, max reasoning at $0.15/M input, $0.03/M cached, $0.50/M output first party, now also on Novita and Runware, and on SCX.ai at $0.13/M input and $0.40/M output.
  • Qwen3.8 Flash (qwen3.8-flash): 1M context and 131K output at $0.15/M input, $0.016/M cached, $0.47/M output on Alibaba Cloud Model Studio's Singapore region, with a Novita deployment alongside.
  • DeepSeek V4 Flash Vision Exp (deepseek/deepseek-v4-flash-vision-exp, beta): image input for DeepSeek's V4 Flash line with time-based pricing, reasoning, tools, and JSON output.
  • Consensus Protocol joins as a provider, serving DeepSeek V4 Flash over an OpenAI-compatible API at $0.14/M input and $0.28/M output with a 524K context.

Stable Image Model IDs

gemini-3-pro-image and gemini-3.1-flash-image are now the canonical IDs for Google's image models, matching the stable upstream releases. The -preview IDs keep working as aliases with identical provider mappings, so nothing changes for existing callers.

Browse all models → | Providers →