Context-window validation, 2026-10-02

For the latest measured model capacity, latency and cost, see model performance. This page retains the October 2 client-profile validation and its configuration-specific limits.

These synthetic text probes ran through the authenticated local mayu gateway, with one configured target per public model. They do not qualify arbitrary tools/history, model quality, another deployment/region, or an exact maximum. No prompt, virtual key, account identifier or session identifier is recorded in the machine-readable evidence.

Public model	Declared bound	Observed accepted input	Result
openai.gpt-6.1-sol	1,050,000	600,026	About 1,048,000 synthetic units returned an unclassified upstream 400; maximum unresolved.
qwen.qwen3-coder-next	256,000	254,029	Accepted; client usable window may increase from 204,800 to 243,200.
global.xai.grok-4.7	500,000	430,037	Accepted; client usable window may increase from 400,000 to 425,000. Maximum unresolved.

Sol 6.1 also rejected reasoning.effort: "none" on a short request. The same short request succeeded without that native option. Native probes therefore leave effort unset; bridge probes use the existing neutral sentinel. The long native request still returned 400 without the effort option. The gateway hides the upstream error body, so neither error is promoted to an exact context limit. Sol 6.1's existing client bound/headroom remains unchanged. Its operational native declaration differs from the separate Converse example/worktree.

Installed Codex 0.159.3 runtime tests against loopback fakes confirmed model switching resolves Grok to 425,000 usable tokens and Qwen to 243,200, even with an inherited model_context_window=1048576. The local installed launcher and icodex shell integration were updated after validation; a new shell/client loads the operator profile. The gateway and global Codex config were not changed.

The total campaign ceiling was $30. Configured-price estimates for successful usage and conservative reservations for unknown failures are kept separately in the artifact; they are not AWS bill reconciliation. Failed calls are not assumed free. Remaining budget was insufficient for the next near-window Astra/Sol 6 probe, so those models remain unqualified rather than guessed.

To use the reviewed Qwen/Grok client adjustments:

export INFERPLANE_CODEX_CONTEXT_PROFILES="$PWD/examples/codex/context-profiles.json"
icodex -m qwen.qwen3-coder-next

The launcher reads the profile at startup, keeps the gateway's declared cap, and changes only usable/compaction percentages. For Qwen these are 95%/90%; for Grok they are 85%/82%. A changed gateway bound refuses the stale profile. Restart the client after installing a profile. Explicit caller compaction settings can still affect runtime compaction.

The reusable probe prints a plan unless --execute is supplied:

python3 scripts/inferplane-context-probe \
  --config /path/to/gateway-config.json --model qwen.qwen3-coder-next \
  --units 254000 --max-spend-usd 30 --ledger /path/to/campaign.json

For execution, privately export INFERPLANE_API_KEY and append --execute. Use the same ledger for every model/call in a campaign. The command locks the ledger across calls, persists the worst input/cache-price reservation before sending, never retries, bypasses ambient proxies, and refuses redirects. Prices come from local overrides and must conservatively cover the deployment's actual tiers; hosted tools/search are not invoked. No provider-secret references are resolved. Unknown failures keep their full reservation. A success is recorded as a qualified observation only when the public model remains unchanged and no model-substitution header is present. This is operator evidence, not financial authority or a proof against an adversarial/stale backend declaration.