A generated title card reading 82% is a tweet, not a leaderboard entry, with the subtitle GPT-6.2 Sol, two days after launch, and three chips giving the Browser Use post of 18 November at 82%, the public board of 29 October at GPT-6.1 Sol 81.3, and the shared rate card of $2 in and $10 out.
Guides & Insights

GPT-6.2 Sol's 82% Browser-Use Score Is a Tweet, Not a Leaderboard Entry

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

On 18 November, Browser Use posted that GPT-6.2 Sol is the top model on its browser-agent benchmark: 82%, up from 81% for GPT-6.1 Sol, 35% cheaper per task, twice as fast, and 36% fewer steps. Greg Brockman amplified it the same day with "GPT-6.2 is incredible at browser use." Two days earlier, O​penAI had shipped GPT-6.2 Sol at $2 per million input tokens and $10 per million output. The claim is a good one and it is probably directionally right. It is also, as of today, the only place those four numbers exist. Browser Use's own public leaderboard still lists GPT-6.1 Sol at 81.3 and has no GPT-6.2 Sol row at all — its results page carries a last-modified stamp of 29 October 2026, and the URL that would hold a per-model page for GPT-6.2 Sol returns a 404. So the honest version of this story is narrower and more useful than the tweet: a benchmark vendor is telling you the model they now rank first is better, cheaper and faster than the model they ranked first three weeks ago, and the board you would normally check to confirm that has not caught up.

What Browser Use actually posted, in full

The quoted post came from Browser Use's own account at 00:09 UTC on 18 November. It reads: "GPT-6.2 Sol is the top model for browser use 👀 — 82%, up from 81% for GPT-6.1 Sol; 35% cheaper; 2x faster: 12 min instead of 22 min; 36% fewer steps." Those are four separate claims of very different strength, and they should be separated before any of them is used to justify a migration.

• Score — 82 against 81. One point on a 0–100 mean weighted rubric score, on a benchmark whose published methodology describes 180 tasks run on live websites. One point is within the noise band of a benchmark that re-runs against changing third-party sites.

• Cost per task — 35% cheaper. This is the strongest claim in the set, because it is the one an agent operator can reproduce cheaply and because it is arithmetically constrained: the token prices did not change, so the saving has to come from somewhere else.

• Wall-clock time — 12 minutes instead of 22. This is the claim that contradicts the vendor's own published data most directly, and it is dealt with below.

• Steps — 36% fewer. In a browser agent, steps are tool calls, and tool calls are where an agent loop spends both money and time. A 36% reduction in steps plausibly explains most of the other three numbers at once.

Browser Use Bench v2.1 is not a marketing exercise. Its methodology card is explicit that the score is a mean weighted rubric score per task on a 0–100 scale, that the cost column is estimated model cost per task in USD and excludes search, judge, browser and runner costs, and that the runs are on 180 live websites rather than a frozen test set. That is a real measurement regime. It is also a regime where a single-point score move and a live-site re-run are not the same kind of evidence, which is why the board has historically been more conservative than its own social posts.

The four numbers are not on the board

We checked the leaderboard at browser-use.com/benchmarks/agents on 19 November 2026. The structured dataset metadata embedded in that page still reads "Browser Use Bench v2.1 model results" with a dateModified of 2026-10-29. Searching the page source for the string "GPT-6.2" returns zero matches. The per-model page at browser-use.com/benchmarks/models/gpt-6-2-sol does not exist — it serves a 404 stub. And the page's own FAQ still describes the standing order as "GPT-6.1 Sol high: 81.3 of 100 at $1.92 a task, on 180 tasks on live websites. Next is GPT-6 Sol high at 79.4."

A generated two-column scoreboard comparing the Browser Use post of 18 November against the public leaderboard of 29 October: score 82 versus GPT-6.1 Sol 81.3, cost per task 35% cheaper versus $1.92, time per task 12 minutes versus 14.3 minutes, steps 36% fewer versus not published, input $2.00 on both, output $10.00 on both.

That is not an accusation of bad faith. Benchmarks are published on a cadence and announcements are not; the account that posted on 18 November is the one that runs the harness, so it has the numbers before the site does. But it does define the epistemic status of this piece, and of any procurement decision made from it this week: the 82%, 35%, 2x and 36% figures are vendor-reported and unpublished, in the specific sense that the vendor's own published table does not contain them yet. A team that moves a production browser agent to GPT-6.2 Sol today is acting on a benchmark announcement, not a benchmark result.

The good news is that the announcement is cheap to falsify. If the numbers hold, an updated board with a GPT-6.2 Sol row should appear within days, and the four figures should survive contact with the JSON. If the row appears with different numbers, you will know before you have rebuilt anything.

The baseline in the tweet does not match the baseline on the board

Here is the detail worth flagging, because it is the kind of thing that gets lost when a claim travels as a screenshot. The tweet gives GPT-6.1 Sol's previous result as "81%" and "22 min". Browser Use's published leaderboard gives GPT-6.1 Sol at high reasoning effort as 81.3 of 100 at $1.92 per task, with a median of 14.3 minutes. The score is close enough — 81 and 81.3 are the same number rounded differently. The time is not. Fourteen minutes and twenty-two minutes are not the same measurement.

A capture of the Browser Use Bench leaderboard on 19 November 2026 showing 17 models on 180 live-website tasks, GPT-6.1 Sol high leading at 81.3 at $1.92 per task and a 14.3 minute median, GPT-6.1 Terra high at 79.0 for 35% less, and a last-updated stamp of Oct 29, 2026, with no GPT-6.2 Sol row visible.

There are innocent explanations. The leaderboard's time column is a median across 180 live tasks, and medians are sensitive to which tasks get dropped for a site failing; 22 minutes could be a mean, could be a high-effort-only subset, could be a different task slice, or could be the runtime before a harness fix landed between 29 October and 18 November. There is also a less innocent one: the tweet's time figure is a comparison chosen to make the new model look faster, and the honest denominator for a 2x speed claim is whatever the same harness produced for the old model on the same day.

You cannot resolve that from outside. What you can do is refuse to import the ratio. "2x faster" here means "the vendor's 18 November harness measured 12 minutes where it measured 22 minutes for the previous model," not "GPT-6.2 Sol is twice as fast as the model the public leaderboard says runs in 14.3 minutes." If speed is the reason you are considering the switch, the number to plan against is the 14.3-minute published figure until a new board replaces it, because that is the only one with a methodology page behind it.

Why 35% cheaper is the claim with arithmetic under it

The pricing makes the cost claim checkable in a way the others are not. GPT-6.2 Sol and GPT-6.1 Sol carry the same rate card, per O​penAI's own model pages:

• Input — $2.00 per million tokens on both.

• Cached input — $0.10 per million tokens on both. Cache writes are $2.50 per million for a 30-minute cache, $5.00 for the 12-hour write, on both.

• Output — $10.00 per million tokens on both.

• Long context — above 272,000 input tokens, input and cache rates double and the output rate goes to 1.5x for the whole request, on both models. The cliff sits in the same place.

• Batch — 50% of standard rates on both; Fast mode doubles, Ultrafast triples through 28 February 2027 and sextuples from 1 March 2027 on both.

• Context — 1,050,000-token window, 922,000 max input, 128,000 max output, text and image in, text out, on both. GPT-6.2 Sol's knowledge cutoff is 1 June 2026 against 30 April 2026.

Identical inputs, identical per-token prices. So a 35% drop in cost per task cannot be a price cut. It has to be fewer tokens bought, and the tweet supplies the mechanism: 36% fewer steps. A browser agent's token bill is dominated by conversation history replayed on every step, so steps and cost scale together almost linearly in a long run. Cut the step count by a third and the bill falls by roughly a third, which is exactly the size of the reported saving. The two claims corroborate each other, which is a good sign for both.

It is also a warning about how to read the saving. A 35% cheaper task is not a 35% cheaper month if the cheaper model is being given tasks the old one failed at, or if fewer steps means the agent is terminating earlier and returning partial work. The rubric score of 82 is what rules that out — a run that quits early with a third fewer steps would score lower, not one point higher — but one point is a thin margin to carry that much weight.

Token prices being passed through unchanged is the part of this that is genuinely boring and genuinely useful. On OrcaRouter, the 0% markup model means a provider's list price is the price you pay, so when a vendor does move a rate card — as O​penAI did for GPT-6.2 Astra, which dropped 20% on input and output the same day — the change is live on the platform the day it happens rather than waiting on a reseller to reprice. That cuts both ways too: there is nothing to arbitrage, and no discount to gain by moving. If you are running GPT-6.1 Sol through one API key today, a GPT-6.2 Sol model call would cost the same per token it costs at the vendor.

What matters for an agent loop, beyond the headline

Four operational details from O​penAI's model page matter more to a browser-agent deployment than the benchmark number does, and none of them appear in an 82% headline.

• Reasoning effort — GPT-6.2 Sol rejects none. The lowest setting is a real one, not a switch to turn reasoning off. GPT-6.1 Sol accepts the full low-to-max range. If your harness hardcodes "no reasoning, just act" for latency, that configuration does not port.

• Batch — supported, at 50% of standard rates. Browser-agent trajectories are usually latency-sensitive, so this matters mainly for offline evaluation runs rather than production.

• Endpoints — Responses, Chat Completions and Batch. Realtime, Assistants, Fine-tuning, Images, Videos, Speech and Transcription are all unsupported. Fine-tuning not being available is the one that bites teams that tuned a browser policy on the previous model.

• Tools — computer_use, web_search, mcp, hosted_shell, apply_patch, skills, tool_search, code_interpreter, image_generation and file_search are listed as supported. For a browser agent the interesting pair is computer_use alongside tool_search: a harness that can discover tools mid-run is the plausible source of a third fewer steps, because the agent stops guessing at an action that does not exist and then retrying it.

• Rate limits — 5,000 requests and 1,000,000 tokens per minute at the Build tier, 10,000 and 4,000,000 at Launch, 15,000 and 40,000,000 at Grow.

How to try it without betting the production path on it

The right response to an unpublished benchmark claim is a sample, not a migration. Pick twenty to fifty of your own browser tasks, run them on both models through the same harness on the same day, and compare three numbers: completion rate on your rubric, tokens consumed per completed task, and p50 and p95 wall-clock time. Then compare the two results against each other rather than against Browser Use's 82, because their task distribution is not yours and a one-point gap will not survive the translation.

Two practical notes for that test. First, pin the reasoning effort explicitly on both sides and use a setting both models accept, or you will spend the run measuring a configuration difference. Second, log step counts, since 36% fewer steps is the mechanism behind the cost claim and the first thing to confirm or refute in your own data.

For the trial itself, failover is the cheap way to run an unproven model next to a proven one: route the sample to the new model with automatic fallback to the model you trust, so a refusal, a timeout or a rate-limit spike costs you a retry rather than a failed job. That is useful here specifically because GPT-6.2 Sol is two days old and its failure modes on live sites are exactly what the benchmark is meant to measure and has not yet published. Be aware of the catalogue state, though: OrcaRouter lists 217 models including GPT-6.1 Sol, the model that currently tops the Browser Use board, but GPT-6.2 Sol is not on it yet — for that one you would go through O​penAI's own API directly. The previous model at pass-through pricing is a sound baseline to measure the new one against, and that is the role it should play this week.

A capture of the OrcaRouter model catalogue listing more than 200 routable models in a provider-and-model table, with rows including openai/gpt-6.1-sol, openai/gpt-6-sol, openai/gpt-6-luna and openai/gpt-6-astra, and no GPT-6.2 Sol entry.

What would change this story

Three things, in order of how much they would move the answer. A GPT-6.2 Sol row appearing on browser-use.com/benchmarks/agents with figures matching or near the tweet — that converts a vendor announcement into a vendor result and makes the migration argument much stronger, though it still would not be independent. An independent third-party measurement of browser-task time and cost on the same task set — that is what would settle the 12-versus-22-minute question, and it is the number to watch for. And a published task-level breakdown, because "36% fewer steps" is either a harness-level improvement that generalises to your agent or a quirk of 180 specific websites, and only the per-task data distinguishes those.

Until at least the first of those lands, the defensible summary is this: GPT-6.2 Sol costs the same per token as GPT-6.1 Sol, the small score gain and large cost gain Browser Use reports are internally consistent with a one-third cut in tool calls, and the speed claim rests on a baseline that does not match the vendor's own published median. That is enough to justify a two-day evaluation on your own tasks. It is not enough to justify a rewrite.

Compared in this article1

Detected from this article · Benchmarks: Artificial Analysis · updated daily