The Cost of ASTRA 6: 78.6% Less Token Capacity and 52.0% Less Account Value at API Token Prices

The answer

In the matched weekend windows, one percentage point of weekly quota carried 18.270 million tokens under GPT-5.6 Sol and 3.910 million normalized tokens under GPT-6 Astra. Expressed at the requested resolution, that is 1.827 million versus 0.391 million tokens per 0.1% of weekly quota.

The result is an account-level observational estimate. It shows what changed in this measured subscription workload. It does not prove that every account, geography, or workload received the same capacity change.

The quota curve

The x-axis is quota consumed. The y-axis is locally recorded tokens processed. A steeper curve means the subscription carries more token work before the weekly meter advances.

Why the weekly window feels as if it is burning

The direct measurement is simple: Astra consumed 4.67× more weekly quota per token than the matched Sol baseline. The meter therefore advanced much faster even when the visible work did not look 4.67× larger.

“Energy phase” is a useful operational name for this regime, not an OpenAI telemetry field. The logs do not expose joules, GPU time, scheduler priority, or the private quota formula. What they do expose is the outcome: the same 0.1% of weekly quota carries only 21.4% as many Astra tokens as Sol tokens.

Measured: 4.67× quota burn per token and 78.6% less token capacity. Inference: the weekly meter is compute-weighted by model class or service cost rather than being a raw-token counter. This is why the window feels as though it is burning heavily.

The public API rate supports the direction of that inference. Astra’s published input, cached-input, and output prices are each 2.5× Sol’s. The observed quota penalty per token is larger—4.67× in the direct pair—so the subscription meter is not behaving like a simple token counter or a strictly proportional copy of public API pricing.

Research method

1. Parse the local event corpus.Scan Codex rollout JSONL for token-count events, token-usage records, turn context, session metadata, model identifiers, timestamps, and rate-limit snapshots.
2. Isolate the weekly meter.Keep only limit_id=codex readings with a 10,080-minute window. Five-hour and model-specific secondary meters are excluded.
3. Build a response-level token ledger.Deduplicate by response ID. For older cumulative counters, sum only positive deltas within a monotonic segment. Cross-check that total against per-response records and a naïve cumulative reconstruction.
4. Reconstruct reset windows.Cluster near-identical resets_at timestamps, assign a canonical window, and compute quota movement as max minus min within each window. When a weekend crosses an instant reset, sum both window segments instead of treating the drop to zero as negative use.
5. Match comparable periods.Compare a Sol-dominant weekend beginning August 28 with an Astra-dominant weekend beginning September 4. Both use the same local machine, account, instrumentation, weekend definition, and ledger rules.
6. Normalize the Astra model mix.The Astra weekend contained 304.834 million Astra tokens and 37.995 million Sol tokens. The Sol-attributed contribution is removed using the directly measured Sol quota rate, yielding 312.796 million Astra-normalized tokens across 80 percentage points.

Controls that materially changed the result

Results

The token-capacity effect is larger than the API-value effect because Astra tokens have a higher public list price. Converting each model’s token mix with the published rates produces an estimated full-week API-equivalent value of about $1,143 for Sol and $548 for Astra, a reduction of approximately 52%. This is a list-price equivalence, not an invoice or a claim about OpenAI’s internal cost.

Interpretation and limitations

• The design is observational and uses one account, not a randomized experiment.
• Quota percentage is server-reported at integer precision. The 0.1% curve is interpolation.
• Workload composition, context length, cache behavior, and tool use can affect quota burn.
• Full-week capacities extrapolate measured partial-window rates to 100%.
• The direct 78.6% estimate should be read with the 75.3% pooled estimate and ≈64% conservative sensitivity check.

Conclusion: after the workload moved from Sol to Astra 6, the measured subscription delivered about one-fifth as many tokens per unit of weekly quota. The best direct estimate is a 78.6% reduction in token capacity. The principal aggregation methods place the reduction between 75.3% and 78.6%.

Reproducibility note: all numerical claims above were recomputed from the local rollout corpus with one reset-aware ledger. Session text was not published. Only aggregate counts, model labels, timestamps, quota deltas, and token totals appear in this report.