Best LLM for Browser Agents: Sonnet 5.5 vs Opus 5.5 vs Holo4

A 30-step browser-agent task prices out at about $4.30 on Claude Opus 5.5 — or roughly $0.68 with prompt caching on. On list prices, that gap beats anything you would save by switching models, and it is the arithmetic that should pick your model, not the leaderboard.

A model roundup published 27 October 2026 pulls the current browser-agent numbers from each vendor's own pages and puts them side by side. Its own caveat is worth repeating up front: providers score on different OSWorld versions, subsets and harnesses, so every figure below is that vendor's claim, not a joint ranking.

What the vendors claim

Model	Claimed score	Price / 1M (in / out)
Claude Opus 5.5	81.8% OSWorld 2.1 (partial credit)	$4 / $20
Claude Sonnet 5.5	80.1% OSWorld 2.1	$2 / $10
Claude Haiku 5.5	72.4% OSWorld 2.1 offline subset	$0.10 / $0.50 under 100K prompts
Holo4 35B-A3B (open weights)	80.8% OSWorld, $0.05 per task on H's harness	$0.30 / $2.00
GPT-6 Astra	72.6% OSWorld 2.0 offline subset	$10 / $50
Gemini 4 Argon	69.2% OSWorld 2.0 offline subset	$2 / $10 until 30 Nov 2026

Anthropic shipped Opus 5.5 on 22 September 2026 and put it at 81.8%, ahead of Claude Fable 5.1 at 80.7% — and Fable 5.1 lists at $10 / $50, so Opus 5.5 beats it on this benchmark at 40% of the price. Sonnet 5.5 (28 September) lands 1.7 points behind Opus at half the cost, with cache hits billed at $0.10 per 1M. Haiku 5.5 (7 October) claims 72.4% where Anthropic lists GPT-6 Luna at 48.9%, but its price is tiered: $0.10 / $0.50 up to 100,000-token prompts, then $0.50 / $2.50 above that.

Gemini 4 Argon went GA on 13 October with computer use still in Preview; its model card claims 39.5% on Agent's Last Exam (computer use), above Astra at 34.2% and Opus 5.5 at 38.2%. Its $2 / $10 rate runs only until 30 November 2026, then doubles to $4 / $20. OpenAI publishes no browser or OSWorld score for GPT-6.1 Terra ($1 / $5) or Sol ($2 / $10) on its model pages.

Step count is the real cost driver

A task a human finishes in two minutes can take 30 to 60 agent loops, and every loop resends the conversation. Two fixed costs set the floor on Claude: declaring Anthropic's browser toolset (browser_toolset_20260801, now GA with 31 member tools) adds roughly 6,600 input tokens to every request, and a 1280×800 screenshot bills at one visual token per 28×28-pixel patch — 46 × 29 = 1,334 tokens per image.

Run the roundup's worked example — a 30-step task, each step adding a screenshot plus about 450 tokens of action and tool text — and total input reaches roughly 1.03M tokens. At list prices that is about $4.30 on Opus 5.5, $2.15 on Sonnet 5.5 and $0.11 on Haiku 5.5 uncached; at a 95% cache-hit rate, about $0.68, $0.34 and $0.02. The author flags these as illustrative figures from list pricing, not measured runs, but the direction holds: caching cuts the bill by more than 80%, and trimming old screenshots out of history saves more than downgrading the model.

Picking one

The roundup's advice is to start on Sonnet 5.5 and move up to Opus 5.5 only when success rate on your real tasks is the bottleneck, down to Haiku 5.5 or Gemini 3.8 Flash ($0.75 / $3.75 until 31 December 2026) for short repetitive flows, and to Holo4 — whose 35B-A3B weights are listed as Apache 2.0 — when data cannot leave your infrastructure. One factor the price table misses: Opus 5.5 is reported tied with Fable 5.1 for the lowest prompt-injection attack success rate on Gray Swan's benchmark, which matters when your agent is reading pages you do not control.

Before benchmarking anything, check whether the page your agent clicks through is backed by a JSON API. One API call replaces 30 steps, and the economics above stop applying entirely.