This week brought four new models worth evaluating. Three are text-focused models aimed at agentic and coding work, and one is an image generation and editing model. They differ more on price and intended workload than on headline quality. Three of them score 100 on our quality index, so the real decision comes down to cost, context needs, and the shape of your workload.
StepFun Step 5 Preview: A Sparse Flagship for Agentic Work
Step 5 Preview is StepFun's flagship model for agentic tasks. It uses a sparse Mixture-of-Experts architecture with 27B active parameters out of 600B total. Only a fraction of the network runs per token, which is how StepFun can offer a large model at a moderate price. The vendor positions it as strong in software engineering and professional work.
Pricing is $1.00 per million input tokens and $2.70 per million output tokens, and it scores 100 on our quality index. That makes it the most expensive of the three text models here, though it is still cheap compared with many frontier options. Agent loops generate a lot of output tokens through planning, tool calls, and code edits, so the $2.70 output rate is the number to watch.
Step 5 Preview fits teams building coding agents or multi-step professional workflows who want to test a large MoE model from a vendor outside their current stack. The "Preview" label matters. Expect behavior to shift before a stable release, and pin your evaluations to dated runs.
Claude Haiku 5.5: The Volume Workhorse
Claude Haiku 5.5 is Anthropic's small, fast model for high-volume, cost-sensitive jobs such as summarization, subagents, and browser use. It succeeds Claude Haiku 4.5, and Anthropic reports stronger coding and computer use than its predecessor.
The pricing is the main story. Standard access costs $0.10 per million input tokens and $0.50 per million output tokens. The batch tier halves that to $0.05 input and $0.25 output. Both variants score 100 on our quality index. Haiku 5.5 is the cheapest model in this release group by a wide margin. It costs roughly a tenth of Step 5 Preview on input and under a fifth on output.
Haiku 5.5 makes the most sense as a subagent inside a larger system. An orchestrator can send it document triage, page navigation, extraction, and first-pass summaries. Its computer use improvements make it a reasonable choice for browser automation where many short steps add up. If your workload can tolerate delayed results, such as nightly classification runs, evaluation sweeps, or backfilling summaries across an archive, use the batch tier and cut the bill in half.
Mistral Large 4: Long Context at a Mid-Range Price
Mistral Large 4 is a frontier multimodal model that accepts text and image input. Mistral built it for reasoning, coding, and agentic workloads. Its defining feature is a 1M-token context window.
It costs $0.68 per million input tokens and $2.09 per million output tokens, and it scores 100 on our quality index. That puts it below Step 5 Preview on both input and output price. The input rate matters most here because long-context work is input-heavy. A single prompt that fills a large share of a 1M-token window costs a fraction of what the same prompt would cost on more expensive models.
Choose Mistral Large 4 when your task requires reading a whole codebase, a long contract set, or a large research corpus in one pass. Its image input also helps for workflows that mix screenshots, diagrams, or scanned pages with text. For short, repetitive calls, Haiku 5.5 is the cheaper option.
Google Nano Banana 2.1: Image Generation and Editing
Nano Banana 2.1, also called Gemini Nano Banana 2.1, is Google's image generation and editing model on the Flash tier. It succeeds both Nano Banana 2 and Nano Banana Pro. Google cites improvements in product recontextualization, which means placing a product into new scenes and settings.
Pricing is $1.50 per million input tokens and $7.50 per million output tokens, and it scores 83.7 on our quality index. That score measures a different job than the text models do, so do not compare it directly with the others. The output price is the highest in this group, which is typical for image output.
Nano Banana 2.1 suits e-commerce and marketing teams that need to show products in varied contexts without new photo shoots. If you currently run Nano Banana 2 or Nano Banana Pro, test 2.1 against your existing prompts before switching, since it replaces both.
How to Choose
| Model | Input $/Mtok | Output $/Mtok | Quality | Best fit |
|---|---|---|---|---|
| Step 5 Preview | 1.00 | 2.70 | 100 | Coding and agentic flagship trials |
| Claude Haiku 5.5 | 0.10 | 0.50 | 100 | Subagents, summarization, browser use |
| Claude Haiku 5.5 (batch) | 0.05 | 0.25 | 100 | Offline high-volume jobs |
| Mistral Large 4 | 0.68 | 2.09 | 100 | 1M-token context, multimodal input |
| Nano Banana 2.1 | 1.50 | 7.50 | 83.7 | Image generation and product editing |
A practical stack for most agent builders pairs Mistral Large 4 or Step 5 Preview as the planner with Claude Haiku 5.5 handling subtasks. Route anything that does not need an immediate answer through Haiku's batch tier. Before committing, run your own task set through each candidate and compare cost per completed task rather than cost per token, since a model that finishes in fewer steps can be cheaper even at a higher rate.