A cross-module bug often takes more time to reproduce, merge and verify than to patch. Carbon targets long-running work and Antigravity now previews /swarm. If you already pay for Claude or ChatGPT, price the bottleneck that keeps your work from shipping.
Google Carbon launch · Carbon and /swarm preview in Antigravity
Choose the surface where you need Carbon
Google announced public Carbon access on November 10, 2026. The personal Gemini Apps table lists Argon for Google AI Pro chat and tasks, and Carbon for Ultra. The model card lists a combined input and output context of up to one million tokens. Check the individual surface for file limits, available features and remaining usage.
Before paying, open the model picker and usage page in the account you will use. Work and school access follows the Workspace licence and administrator settings.
| Surface | Published access | Check before paying |
|---|---|---|
| Gemini Apps | Argon on personal Pro and Carbon on Ultra | Model picker and usage in your account |
| Google Antigravity | Carbon rolling out to Ultra, /swarm preview in app and CLI | Client version, workspace access and usage |
| Gemini Developer API | Paid API access for public Carbon | Project entitlement, billing and December rates |
Google Carbon launch · Gemini Apps personal model access · Google DeepMind Carbon model card
Set a reproducible debugging baseline
Create a sanitized event aggregation case. Tenants A and B both emit event_id=42, then A retries. An older event arrives late and attempts to replace A’s newer state. Pin the commit and test command. The acceptance checks count A and B once each, keep A’s new state and pass the existing single-tenant regression suite. Give every candidate the same spec, input and permissions.
Start with three runs on your current Codex or Claude Code access. Reset the workspace before each attempt and record the actual model, modified files, first-pass acceptance, elapsed time and minutes you spent reviewing or fixing the result. Keep failed runs in the ledger.
{
"dedup": {
"events": [
{
"tenant": "A",
"event_id": 42
},
{
"tenant": "B",
"event_id": 42
},
{
"tenant": "A",
"event_id": 42
}
],
"expected_unique_counts": {
"A": 1,
"B": 1
}
},
"state_order": {
"events": [
{
"tenant": "A",
"event_id": 43,
"version": 2,
"status": "delivered"
},
{
"tenant": "A",
"event_id": 41,
"version": 1,
"status": "in_transit"
}
],
"expected_status": {
"A": "delivered"
}
}
}Use the supplied commit as the baseline. Reproduce the shared event_id and retry issue across A and B. Verify that an old event cannot overwrite a newer state. Return the smallest patch, failing and passing test logs, and open questions. Run tests only in the sample workspace and pause before external actions.run_id,commit,model,surface,accepted,elapsed_min,review_min,rework_min,api_cost_usd,quota_event
OpenAI Codex product page · Anthropic Claude Code product page
Assign parallel work and agree on the merge checks
Google’s /swarm preview has a coordinator, workers in separate git worktrees and verifiers. You can assign models by role. Give one worker the deduplication key and another the state update. Have the verifier test the combined diff and the coordinator return the patch with its logs. Use a model and workspace you can already access.
Check both patches for changes to shared types and migration scripts before merging. Keep the input commit, diff, exit status and failure logs. When both edits depend on a shared interface, agree on that interface first. Google also warns that parallel teams consume quota faster than normal chats.
- Verify model access
- Pin the fixture and acceptance checks
- Record the single-agent baseline
- Compare parallel results and review effort
- Price the purchase using live quotes
Calculate API usage and subscription costs separately
The published standard developer API rates through November 30 are $2 per million input tokens and $10 per million output tokens. From December 1 they become $4 and $20. Output includes thinking tokens. At 80,000 uncached input and 20,000 billable output tokens, the model charge would be $0.36 and $0.72 respectively. Tool charges, cache use and retries can change the total. App usage follows its own limits.
Suppose you accept four jobs per month on either route. Human review and rework fall from 55 to 25 minutes each at $60 per hour, saving $120 per month. If the second plan and incremental costs add up to $90, the estimated saving is $30 and three jobs cover that extra spend. Replace these illustrative figures with your own quotes, usage and rerun costs.
Turn the results into a purchase decision
Keep your current plan when it clears the acceptance bar on time. If a sound workflow stalls only at usage peaks, compare the current Claude and ChatGPT tiers or Codex credit packs for an eligible account. If repeat tests show that Carbon resolves cross-module failures or reduces review effort, check eligibility and your local price directly with Google.
On ClaudeMax, open the relevant product card to confirm availability, renewal and whether it tops up your account or delivers a separate one. Check top-up eligibility if you need to preserve existing projects, history or workspace access. ClaudeMax lists Claude and ChatGPT options. Carbon access comes from your Google account or API project.