GPT-6 Astra vs Opus 5 vs Muse Spark 1.3: Computer Use
The Short Answer
As of September 5, 2026, three models are credibly competing for autonomous computer use — driving a real desktop or browser through a perception-action loop:
| Model | OSWorld 2.0 | Price /MTok | Released |
|---|---|---|---|
| GPT-6 Astra | 72.6% | $10 / $50 | Sep 3, 2026 |
| Claude Opus 5 | 70.2–70.6% | $5 / $25 | Jul 24, 2026 |
| Muse Spark 1.3 | 66.9% | $1.25 / $4.25 | Sep 2, 2026 |
| Gemini 3.8 Flash | 59.0% | $0.75 / $3.75 | Sep 2, 2026 |
| GPT-5.6 Sol | 65.7% | $4 / $20 | — |
The honest summary: Astra wins on speed far more decisively than it wins on accuracy.
What Actually Changed on September 3
OpenAI released GPT-6 Astra on September 3, 2026, and made computer use the headline capability rather than a footnote. The number OpenAI published is 72.6% on OSWorld 2.0 at roughly 40 minutes per task, compared with GPT-5.6 Sol at 65.7% at roughly 75 minutes — described as about 47% less time per task.
That latency figure is the interesting one. A seven-point accuracy gain is a normal generational step. Halving wall-clock time per task is a different kind of change, because computer-use agents fail on timeouts and drift at least as often as they fail on reasoning. An agent that finishes in 40 minutes gets fewer chances to lose its place.
Astra runs a 1,050,000-token context with 128K max output and a knowledge cutoff of April 30, 2026. Access is staged: enterprise Trusted Access first, then the API, ChatGPT plans and AWS.
⚠️ The pricing trap: prompts above 272,000 tokens reprice at roughly 2x input (~$20/MTok). Computer-use agents accumulate screenshots and scrollback fast, so a long session can silently cross that threshold mid-task. Compact aggressively or cap the loop.
The Three Models Compared
GPT-6 Astra — fastest, most expensive
$10 / $50 per MTok · cached input $1.00 · 72.6% OSWorld 2.0
Beyond computer use, Astra saturates several suites: 97.6% on FrontierMath Tier 4 and 100% on ExploitBench. On abstract reasoning, OpenAI’s announcement table lists 99.9% on ARC-AGI-3 against 30.2% for Opus 5 and 7.8% for Sol — but ARC Prize’s own official standard harness scores Astra at 62.7%, still a decisive lead over Opus 5’s 30.2%. The 37-point spread between those two figures is a good reminder of how much harness choice moves a headline.
Limits: no fine-tuning and no realtime audio at launch, the >272K repricing cliff, and a staged rollout that means “available” depends on your account tier.
Claude Opus 5 — half the price, two points behind
$5 / $25 per MTok · 70.2–70.6% OSWorld 2.0 · 1M context
Opus 5 has been shipping since July 24, 2026 and is now Anthropic’s default Opus. For computer use specifically it remains the sane default: it is half Astra’s token price, it is generally available rather than staged, and the accuracy gap is roughly two points on a benchmark whose error bars are wider than that.
Where it clearly leads is long-horizon agent work. On Terminal-Bench 4.0 — the harder, longer-running agent suite — Opus 5 scores 51.8% against Gemini 3.8 Flash’s 19.1%. Tasks that run for hours rather than minutes are still Anthropic’s strongest ground.
Limits: slower per task than Astra, and on the Artificial Analysis coding-agent harness it costs $8.17 per task against Astra’s $3.27 at xhigh effort — the per-token discount inverts once you measure real token consumption.
Muse Spark 1.3 — cheapest, with an asterisk
$1.25 / $4.25 per MTok · 66.9% OSWorld 2.0 · 1M context
Meta’s September 2, 2026 release is genuinely strong on adjacent work: 88.8% on Terminal-Bench 2.1, 75.4% on DeepSWE v1.1 (ahead of Opus 5 at 74.0), and 98.5% / 98.1% on the MRCR long-context bands. Meta also measured roughly 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2, which matters more than list price in a loop-heavy workload.
⚠️ The asterisk: Meta’s API lists Muse Spark 1.3 as text-input only. A model that cannot accept images cannot do screenshot-driven GUI control in the usual sense, yet trackers report an OSWorld figure for it. The plausible reading is an accessibility-tree or DOM-text harness. Verify the harness before you budget on that 66.9%.
Cost Per Task Beats Price Per Token
This is the single most important correction to make when comparing computer-use models in 2026. Measured on the Artificial Analysis coding-agent harness in September 2026:
| Model · effort | Cost per task | Index |
|---|---|---|
| GPT-5.6 Luna (max) | $0.29 | 57.1 |
| GPT-6 Astra (low) | $1.41 | 62.6 |
| Muse Spark 1.3 (xhigh) | $1.72 | 64.2 |
| Gemini 3.8 Flash (high) | $2.04 | 61.1 |
| GPT-6 Astra (xhigh) | $3.27 | 67.0 |
| GPT-5.6 Sol (max) | $5.00 | 65.0 |
| Claude Opus 5 (xhigh) | $8.17 | 68.1 |
Astra’s per-task output is roughly 2,200–14,000 tokens depending on effort. Gemini 3.8 Flash averages about 48,000. That 7.7x efficiency gap erases a 13x token-price advantage — which is why the “cheap” model is frequently the expensive one on the invoice.
Which Should You Pick?
- Latency-sensitive desktop automation → GPT-6 Astra, low or medium effort. The 40-minutes-per-task figure is the product.
- Cost-controlled production agents → Claude Opus 5, unless you have measured your own per-task token consumption and found Astra cheaper (you may well have).
- Long-horizon multi-hour agents → Opus 5, on the Terminal-Bench 4.0 evidence.
- High-volume scoped tasks → Muse Spark 1.3 or Gemini 3.8 Flash, after confirming the input modality your harness needs.
- Anything touching credentials or payments → none of them unsupervised. Prompt injection through page content remains unsolved; keep an approval gate on state-changing actions.
Last verified: September 5, 2026. Model pricing and leaderboard positions in this category have changed multiple times per month through 2026 — re-check before committing a budget.