I use Astra and Opus side by side, picked by task. The rest run as subagents. The biggest difference is speed. On a subscription, Astra and Sol take ages to finish the work.
Bijan Bowen put it in two lines on X, October 25, 2026. GPT-6.x models = left brain. Claude 5.5 models = right brain. The metaphor is Bijan's. It isn't science. The split below is mine.
The split
| Model | Role | What I give it | The catch |
|---|---|---|---|
| GPT-6 Astra | Side by side with Opus | Coding, code review and debugging, data analytics, math, plans and specs, long runs that stay on the plan, orchestration even better than Opus | Painfully slow. Needs a detailed prompt, then it can beat Opus. |
| GPT-6.1 Sol | Subagent only | Computer and browser use, read-only MCP calls, structured output | Takes ages on a subscription. |
| GPT-6.1 Terra on Fast | Subagent only | Gathering and preparing data, finding things in a big repo | |
| Claude Opus 5.5 | Side by side with Astra | UX/UI and design, writing, coding where the code looks more human, repo-wide refactors, orchestration | Gets your goal from a poor or wrong prompt. Orchestration is almost as good as Astra's, and much faster. |
| Claude Sonnet 5.5 | A very strong subagent | The same subagent jobs as Sol and Terra, and design | Computer use is weaker than Sol. Design is clearly weaker than Opus. |
Model names are from OpenAI's guide and Anthropic's overview. The ratings are my experience as of October 2026, not a benchmark.
This is routing, not a capability list. OpenAI pitches Terra for "visual reasoning, precise computer use, and clear, natural writing." I keep computer use on Sol. Read-only comes from the tools and permissions you allow, not from the model.
Speed depends on what you measure
My subscription experience has a counterpoint. Vals AI's archived Terminal-Bench 2.1 results, updated October 1, 2026, list Astra at 4m34s per test and Opus 5.5 at 6m23s. Both used Terminus 2. That reverses my speed ranking, but measures terminal tasks rather than my subscription sessions.
Fast mode is separate from the model and its reasoning effort. Record all three when comparing speed.
When I prefer one tailored agent
For building several features, one tailored agent beats an orchestrator with subagents. That's my judgment, not a rule for every task.
Anthropic's Diane Sarkis reports a similar enterprise pattern: one agent with skills often improved quality, cost and latency. One customer saw roughly 2x faster runtime. That isn't a controlled comparison of feature development.
Subagents hand back summaries. Each handoff can lose detail. Anthropic still uses subagents for verification, parallel work and separate permissions.
I still orchestrate when the tasks are related, like gathering info or preparing data. There it's much more efficient.
Test it on your own tasks
Anthropic suggests starting with 10 golden examples.
- Pick that many of your real tasks. Run each through one single-agent design and one orchestrated design.
- Hold the model, effort, client and speed tier fixed while comparing designs.
- Run each task more than once to catch inconsistent results.
- Check which tasks flipped, not only the winner. Add wall time and total tokens, so the speed gap has a number.
Log one line per run:
task,design,model,effort,client,speed_tier,run,completed,wall_time,total_tokens
The maintained article is on argbe.tech.
Top comments (0)