Aeon AI Risk Management
Model Economics Lab
Compare model capability, task cost and token usage before choosing a deployment route. Start with current releases and the retained shortlist, then inspect native metrics, sources and evidence tiers. Benchmark cost is not your cost per accepted production outcome.
From benchmark to buying decision
Include failed attempts, human review and rework in cost per accepted task. Compare hosted, private and hybrid routes on the same accepted workload. The linked methods are evaluation guidance, not new benchmark results.
Executive finding
November 6 separates three buying decisions: Inkling 2 Small's hosted launch, Google's November 2 Gemini 2.5 migration notice, and DeepSeek Search/Fetch billing effective October 31 at 16:00 UTC. Hosted availability, mandatory migration and tool charges change different parts of the operating model; none establishes a universal winner or measured client savings.
Latest release evidence
Inkling 2 Small launched November 5 through Tinker with closed weights under current Tinker terms. Apache 2.0 weights are intended next week after external safety testing, not available under that licence today. Artificial Analysis reports Intelligence Index v5.0.1 score 31, $1.18 weighted cost per Intelligence Index task and approximately 150k output tokens per task. The new provisional diamond keeps that native triplet together, not 31 percent coding correctness. Its effort setting is not reported in the reviewed evidence.
Current access and tool costs
Tinker lists serverless inference beta rates of $0.30 input / $0.03 cached input / $1 output per million tokens and a production waitlist. Kimi K3.1 Pro has public model files under a custom licence, with vLLM serving support documented November 4; neither weights nor runtime support prove cheap self-hosting. DeepSeek announced Search/Fetch on October 17 and its server-side tool billing starts October 31 at 16:00 UTC: each endpoint costs $0.75 per 1,000 successful requests, with model tokens additional. Failed and empty requests are not billed. These existing-model access and pricing changes cause no default retirement.
Gemini migration deadline
Google's November 2 notice removes Gemini 2.5 Pro and Flash from AI Studio selection. Starting December 1, API calls to gemini-2.5-pro automatically route to gemini-3.8-flash and calls to gemini-2.5-flash to gemini-3.5-flash-lite, at replacement-model rates. Breaking changes include function-calling/thought-signature validation and thinking-level controls. Inventory aliases, choose explicit targets and regression-test the application and full bill before that date. This is an existing-model routing change, not new launch coordinates or a guaranteed saving.
Access and licence
Carbon's introductory input/output rates are $2/$10 per million tokens, versus standard $4/$20; the expiry was not established. Do not treat restricted access as a production entitlement. Mistral's Ecosystem Open-Weight licence is modified MIT, with a separate commercial-licence requirement for Model-as-a-Service businesses and stated exceptions. Open weights do not mean unrestricted licensing or free infrastructure.
Retained October 23 Kimi evidence
No independent K3.1 Pro coding-agent task coordinate was verified in the October 23 review. Its retained diamond reuses historical Kimi K3 v1.3 score, cost and tokens as a predecessor proxy, not a K3.1 Pro measurement or forecast. The separate launch Intelligence Index was not used for that proxy. API pricing was $2 input / $10 output per million tokens, with a separate $4/$20 high-speed route; weights were planned by November 4, not available as of that review.
Retained October 1 evidence
Sol 6.1 XHigh scores 62.91 at $1.04 per evaluated task; Medium scores 61.41 at $0.70. Argon High scores 63.76 at $5.84 in a different native agent harness. These AA v1.5 observations retain their dates. Argon's restricted initial access and introductory pricing remain part of that historical source context; its later October 13 general availability is a separate availability update, not a rewrite of the plotted observation. Small score gaps are not established superiority.
Subscription efficiency
Anthropic announced higher five-hour limits on eligible subscriptions separately from estimated typical-workload cost savings. No universal subscription-throughput multiplier is established. Maximum-effort coding economics need separate measurement.
Frontier economics
DeepSWE v1.1 gives Astra and Gemini 3.8 Flash the same 74 percent point estimate. Gemini averages $2.36 per task versus Astra's $6.52, while Astra uses fewer output tokens.
DeepSWE v1.1 common-harness view
The independent DeepSWE section retains pass rate, average task cost, output tokens, and agent steps from one mini-swe-agent harness. Use it for direct comparisons among current releases and measured shortlist anchors.
CursorBench 3.2 historical cross-check
This retained snapshot predates CursorBench 4.0. Do not directly compare percentages across versions or treat these older rows as the current shortlist.
Master plot
The horizontal axis is published or explicitly proxied task cost, the vertical axis is each point's published native benchmark score, and marker area represents published or proxied tokens. Intelligence Index v5.0 and v5.0.1, Coding Agent Index and DeepSWE are different measures and must not be ranked together. Filled circles retain comparable dated measurements, diamonds are provisional cross-suite or proxy points, and hollow circles are modeled.
Reader views
Current releases, Popular models, Value challengers, and All / history separate the 19-model default from 43 retained operating points across 26 model families: 14 measured, four modeled and 25 provisional. Every incoming model appears by default. Inkling is a new provider, so no other provider is removed to impose a global cap. November 6 pricing/access-only updates cause no retirement. All historical observations, Kimi K3 and the dated K3.1 Pro predecessor proxy remain unchanged. Users can expand efforts, filter evidence, toggle trajectories and change cost scale.
Historical v1.3 measured anchors
Opus 5 xhigh reaches 66.74 at $8.23 per task. GPT-5.6 Sol max reaches 66.57 at $7.08. Grok 4.5 high reaches 64.44 at $2.59. Kimi K3 reaches 61.34 at $3.18. These are retained observations, not a current-release ranking.
Effort economics
Higher effort creates nonlinear token and cost growth. The last capability points can be valuable for ambiguous, long-horizon, or high-consequence work, but are usually inefficient for routine coding.
Modeled points
Where no comparable public run exists, Aeon uses neighboring model-family effort curves or a vendor-published effort curve. Every estimate is marked as modeled and is not presented as a benchmark result.
Limitations
Output tokens do not include every input, cache, retry, tool, and harness cost. Coding-agent token use can vary substantially by task and harness, so organizations should validate routing against their own repositories and acceptance tests.
Downloadable dataset
The full operating-point dataset is available in CSV and JSON with model, provider, effort, published score, native score metric, task cost and basis, tokens and token metric, evidence status, modeling basis, and source URL.
Dataset license and attribution
Aeon publishes this research compilation for evaluation and citation. Cite Aeon and link to this page when referencing the compilation. Underlying benchmark data and source materials remain subject to their publishers' terms. This notice does not relicense third-party material. Contact info@airiskmanagement.ca for redistribution or commercial reuse.
Update policy
Aeon adds every incoming model to the default plot and removes up to one oldest pre-existing default from the same provider family for each incoming model, using verified release chronology. No other provider is evicted to impose a global cap, no incoming model is removed and pricing/access-only news causes no retirement. All displaced observations remain in All / history.
Operating recommendation
Use a routing ladder that assigns cheaper effort to bounded tasks and reserves frontier settings for ambiguity, long dependency chains, and changes with high failure cost.