Vercel AI Gateway's 8-Model Week: Sonnet 4.6 to Opus 4.7

Vercel's AI Gateway shipped eight changes in about a week: three new coding models, an image model, a new API surface, and a compliance feature with a hard price tag. Taken one at a time, none of it looks urgent. Taken together, it's a routing decision — which model gets which job, and what that costs.

The coding models: Sonnet 4.6, Opus 4.7, and Kimi K2.6

Claude Sonnet 4.6 landed on AI Gateway first, and it's the model most teams will hit by default. Vercel's changelog says the model “is now available on AI Gateway with the 1M token context window” and that it “approaches Opus-level intelligence with strong improvements in agentic coding, code review, frontend UI quality, and computer use accuracy” — Vercel's own framing, not an independent benchmark.

It's also worth noting the mechanics: Sonnet 4.6 proactively executes tasks, delegates to subagents, and parallelizes tool calls, with MCP support for scaled tool use. As a hybrid reasoning model, it delivers both near-instant responses and extended thinking within the same model.

// Illustrative: routing to Sonnet 4.6 via the AI SDK
import { generateText } from 'ai';

const result = await generateText({
  model: 'anthropic/claude-sonnet-4.6',
  prompt: 'Refactor this function for clarity.',
});

Opus 4.7 targets a different job. Vercel describes it as “optimized for long-running, asynchronous agents” that “handles complex, multi-step tasks with reliable agentic execution,” adding that it “shows gains on knowledge-worker tasks, particularly where it needs to visually verify its own outputs” — again, a vendor claim rather than a published benchmark. Vercel also says the model “now has improved memory, with agents that maintain structured memory store across turns seeing more reliable recall and fewer dropped facts without additional prompting.”

The concrete addition here is task budgets: a total token budget set for an agentic turn through a taskBudget parameter. The model sees a countdown of remaining tokens and uses it to prioritize work, plan ahead, and wind down gracefully as the budget runs out — useful for anyone who has watched a long agent loop burn through cost with no exit ramp.

// Illustrative: capping an agentic turn with taskBudget
const result = await generateText({
  model: 'anthropic/claude-opus-4.7',
  taskBudget: 50000, // total tokens for this turn
  prompt: 'Investigate and fix the failing checkout tests.',
});

Kimi K2.6 from Moonshot AI rounds out the coding tier. It “focuses on long-horizon coding tasks, with generalization across languages such as Rust, Go, and Python and across front-end, devops, and performance optimization work,” per its Gateway listing. Vercel also says K2.6 “can turn simple prompts into complete front-end interfaces with structured layouts” and that for continuous, autonomous agents it “improves on API interpretation, long-running stability, and safety awareness during extended research tasks” — again, Vercel's characterization rather than a third-party test.

The efficiency tier: GPT-5.5, GPT-5.4 Mini and Nano

GPT-5.5 ships in two variants. Vercel says “there are 2 variants: GPT-5.5 and GPT-5.5 Pro” and that “both models are tuned for long-running agentic work across coding, computer use, knowledge work, and scientific research, and are more token-efficient than the previous generation.” The Pro variant sits further up the stack: Vercel says it's “built for demanding, multi-step work where response quality matters more than latency,” with “early testing” showing gains in “business, legal, education, data science, and technical research workflows that involve critiquing work over multiple passes and stress-testing arguments” — Vercel's characterization, not a published benchmark.

GPT-5.4 Mini and Nano target a cheaper lane. Vercel describes both as delivering “state-of-the-art performance for their size class in coding and computer use” and says they're “built for sub-agent workflows where multiple smaller models coordinate on parts of a larger task.” Nano specifically “performs close to GPT-5.4 Mini in evaluations at a lower price point,” which Vercel positions for “high-volume use cases like sub-agent workflows where cost scales with the number of parallel calls” — the fan-out pattern where one orchestrator model calls dozens of cheap workers.

New surfaces: GPT Image 2 and the Responses API

GPT Image 2 is the catalog's one non-text model. Vercel says it “can render fine-grained elements including small text, iconography, UI elements, dense compositions, and subtle stylistic constraints, at up to 2K resolution” — a claim worth testing against your own UI mockups or icon sets before committing a pipeline to it.

The Responses API addition is smaller in scope but removes friction for teams already standardized on OpenAI's SDK. As Vercel puts it: “Point your OpenAI SDK to AI Gateway's base URL and use the creator/model names to route requests. TypeScript and Python are both supported.” The underlying functionality isn't new — “all of the functionality in the Responses API was already accessible through AI Gateway via the AI SDK and Chat Completions API” — but teams built around the Responses API no longer need a workaround.

// Illustrative: OpenAI SDK pointed at AI Gateway's base URL
import OpenAI from 'openai';

const client = new OpenAI({
  baseURL: 'https://ai-gateway.vercel.sh/v1',
  apiKey: process.env.AI_GATEWAY_API_KEY,
});

const response = await client.responses.create({
  model: 'openai/gpt-5.5',
  input: 'Summarize this incident report.',
});

Zero Data Retention: the one number that matters

Buried under the model announcements is a pricing change that compliance teams will care about more than any benchmark. Team-wide Zero Data Retention now costs $0.10 per 1,000 requests for Pro and Enterprise plans. Per-request ZDR, by contrast, stays free for Pro and Enterprise, and disallowing prompt training on a per-request basis is free for all users. In other words: request-level privacy controls cost nothing, but a blanket, account-wide guarantee that no request is retained carries a metered price.

Eight changes, one table

Here's the whole catalog in one scan, matched to the job each addition is meant for.

Addition	Best for	Context/cost lever	How to reach it
Claude Sonnet 4.6	Default agentic coding, code review, UI work	1M token context window	AI SDK, model string on Gateway
Claude Opus 4.7	Long-running async agents, multi-step tasks	taskBudget token cap	AI SDK, model string on Gateway
Kimi K2.6	Long-horizon coding across Rust, Go, Python, front-end	Vendor-described stability for continuous agents	AI SDK, model string on Gateway
GPT-5.5 / GPT-5.5 Pro	Long-running agentic work; Pro for quality over speed	Vendor-described token efficiency gain	AI SDK or Responses API
GPT-5.4 Mini / Nano	Sub-agent fan-out, high-volume parallel calls	Nano priced below Mini	AI SDK or Responses API
GPT Image 2	UI mockups, icon sets, dense compositions	Up to 2K resolution	AI SDK image generation call
Responses API support	Teams standardized on OpenAI's Responses API	No shim needed	Point OpenAI SDK base URL at Gateway
Team-wide Zero Data Retention	Compliance teams needing account-wide retention guarantees	$0.10 per 1,000 requests	Pro/Enterprise plan setting

Which one to route to

None of this requires picking one winner — route by task, not hype. Choose Sonnet 4.6 for a single default model across agentic coding, code review, and frontend work. Choose Opus 4.7 when agents run long and unattended: the memory and task-budget additions exist to survive turns that Sonnet-class models weren't built to babysit. Choose Kimi K2.6 if your stack already leans on non-Anthropic, non-OpenAI infrastructure and you want a coding-focused alternative rather than a second frontier bill.

For anything that fans out into dozens of parallel calls, reach for GPT-5.4 Nano before Mini, and save GPT-5.5 Pro for the tasks where a wrong answer costs more than a slow one. Treat GPT Image 2 as a UI-asset tool to test, not a default. And if your compliance team wants a retention guarantee instead of a per-request opt-out, the $0.10-per-1,000-request line item is the one number in this batch that belongs directly in a budget forecast.