← Back to blog
March 15, 2026•Dzianis Vashchuk•6 min read

GPT-5.4 Support Is Live in Vibe — and Kimi-K2.5 Is Now in the Free Tier

Vibe now supports GPT-5.4 for higher-confidence browser work and adds Kimi-K2.5 to the free tier for fast, tool-heavy agentic workflows.

product-updategpt-5.4kimi-k2.5model-routingagentic-web

Today we shipped two model updates in Vibe:

  1. GPT-5.4 support for users who want stronger reasoning and more reliable step selection on messy browser tasks.
  2. Kimi-K2.5 in the free tier so more users can try serious agentic web workflows without immediately paying frontier-model prices.

This is part of the product direction we keep leaning into: the browser operator layer should stay model-flexible.

Why GPT-5.4 matters in Vibe

Some web tasks are not hard because they need one clever answer. They are hard because they need a long chain of correct micro-decisions:

  • which tab or section matters
  • which control is actually interactive
  • when to ask for confirmation instead of guessing
  • how to recover when the page changes shape mid-run

That is where higher-end reasoning models help.

GPT-5.4 is a strong fit for:

  • multi-step browser tasks with ambiguous UI states
  • workflows where a bad click is expensive
  • tasks that benefit from better planning before acting
  • runs where users want fewer retries and less babysitting

In practice, this is the model we expect many users to prefer when the workflow is high-stakes and the cost of a wrong action is higher than the cost of a better model.

Why we added Kimi-K2.5 to the free tier

We also wanted a fast, capable, lower-cost option for agentic browsing.

Kimi-K2.5 stood out because the public technical material points to a model that is unusually focused on tool use, coding, and agentic search workflows, not just static chat.

From Moonshot AI's public materials, Kimi-K2.5 is positioned as a native multimodal agentic model with:

  • a 256K context window
  • an agent swarm mode that can coordinate up to 100 sub-agents
  • support for up to 1,500 tool calls in parallel workflows
  • a claim of up to 4.5x lower execution time versus a single-agent setup on complex tasks

Those are exactly the kinds of traits we care about for browser automation: long context, strong tool behavior, and decent speed under multi-step workloads.

How good is Kimi-K2.5 for agentic web tasks?

Short version: promising enough that it belongs in the product, especially as a free-tier option — and the public BrowseComp data puts it in the top open-weight cluster rather than in a runaway lead.

BrowseComp is the browser benchmark we care about most

As of March 15, 2026, BrowseComp is one of the best public sanity checks for browser-adjacent agent work. The benchmark includes 1,266 questions that require agents to persistently navigate the web and find concise, verifiable answers, rather than just recall memorized facts.

That is why it matters more to us than a generic static QA benchmark when we decide which models belong in Vibe.

BrowseComp leaderboard: GPT-5.4, Kimi-K2.5, and the open-weight field

On the public BrowseComp leaderboard, GPT-5.4 is clearly ahead of Kimi-K2.5. But the more relevant comparison for a free-tier model is where Kimi lands against the other strong open-weight entrants.

Model Public BrowseComp score Weight status What it suggests for Vibe
GPT-5.4 0.827 Proprietary Stronger choice when a browser run is expensive to get wrong.
MiniMax M2.5 0.763 Open-weight Currently the highest-scoring open-weight peer in this set, with an explicit tool-use and search focus.
GLM-5 0.759 Open-weight Slightly ahead of Kimi on public BrowseComp, with positioning centered on agentic engineering.
Kimi-K2.5 0.749 Open-weight Top-tier open-weight result. Moonshot also reports 74.9 with context management and 78.4 in Agent Swarm mode on its own eval setup.
Qwen3.5-397B-A17B 0.690 Open-weight Capable, but materially behind Kimi on this benchmark.

That table changes the story a bit. Kimi-K2.5 is not the single highest-scoring open-weight model on BrowseComp today. But it is firmly in the top cluster: within 0.014 of MiniMax M2.5 and 0.010 of GLM-5, while staying comfortably ahead of Qwen3.5-397B-A17B.

For Vibe, that matters because Kimi's differentiator is not just raw leaderboard placement. Moonshot is explicitly optimizing for long-context, tool-heavy, multi-agent execution. Its public materials emphasize a 256K context window, up to 100 sub-agents, and up to 1,500 tool calls in swarm mode. That is unusually aligned with browser tasks that fan out across many pages, tabs, and search branches.

The broader benchmark story is still strong:

  • On the Kimi-K2.5 model card, Moonshot reports 50.2 on HLE-Full with tools
  • 85.0 on LiveCodeBench
  • 76.8 on SWE-Bench Verified

None of those numbers guarantee perfect browser execution in the wild, but they do suggest Kimi-K2.5 is not just a cheap fallback model. It looks genuinely competitive for tool-heavy tasks and coding-adjacent workflows.

For web tasks specifically, the BrowseComp results are the most interesting signal. They suggest Kimi-K2.5 is worth using when the job looks like:

  • research across many pages
  • extraction plus summarization
  • multi-step browsing with tool calls
  • broad workflow exploration where speed and cost matter

How quick is it?

There are two different meanings of "quick" here:

1) Agent workflow speed

Moonshot's own launch material says Kimi-K2.5 can reduce execution time by up to 4.5x versus a single-agent setup by breaking tasks into a coordinated swarm. If that translates cleanly to real tool stacks, that is a meaningful advantage for web research and multi-page agent flows.

2) Raw serving speed

Serving speed depends on the provider, but public infrastructure partners are clearly optimizing for Kimi. Fireworks claims up to 200 tokens/sec on its Kimi-K2.5 endpoint. That is not a universal model guarantee, but it is a good sign that the ecosystem expects Kimi to be used in latency-sensitive workflows.

Our practical read is:

  • GPT-5.4 is the better default when accuracy and judgment matter most.
  • Kimi-K2.5 is a strong option when you want lower-cost experimentation, fast tool-heavy runs, or a free-tier model that is still meaningfully agentic.

Our current recommendation

If you are choosing between the two inside Vibe today:

  • Use GPT-5.4 for brittle, high-value browser workflows where you want better decision quality.
  • Use Kimi-K2.5 when you want a fast, capable model for research, coding-adjacent browsing, and general agentic web tasks without burning premium credits.

That is exactly why we wanted both in the stack.

Why this matters for Vibe

We do not think the future of browser agents belongs to a single model vendor.

Different workflows want different tradeoffs:

  • best possible reasoning
  • lowest possible latency
  • lower cost at higher volume
  • open-ish model options
  • better coding and tool-use bias

Adding GPT-5.4 and Kimi-K2.5 moves Vibe further toward that model-routing future.

Try it

If you are already using Vibe, you can now:

  • pick GPT-5.4 when the task needs more judgment
  • use Kimi-K2.5 in the free tier
  • compare outcomes on the same workflow and see which tradeoff fits your stack

This is the right way to evaluate browser agents anyway: not by vibes, but by repeated workflows.

References

We will keep expanding model coverage, but the core product principle stays the same: users should be able to choose the right model for the workflow, not the model a browser product picked for them.

Read next

Related posts