This week's model news gives buyers three different decisions: whether to route suitable work to a cheaper API model, whether a stronger frontier model is actually available to them, and whether downloadable weights justify taking on deployment responsibility. Treating all three as a single leaderboard contest misses the buying question.
Our October 30 review selects three previously uncovered developments from October 24-30: GPT-6.1 Terra, Gemini 4 Carbon and the improved open-weight release of Mistral Large 4. No fourteen-day extension was needed. The Model Economics Lab now includes all three in its default view, with their native evidence and limitations visible. These are research observations and evaluation candidates, not measured Aeon client outcomes.
1. GPT-6.1 Terra: evaluate the routing opportunity
OpenAI's October 26 API changelog records GPT-6.1 Terra's release. Its model documentation lists standard prices of USD $1 per million input tokens and $5 per million output tokens, with $0.05 cached input. It supports text and image inputs, and tool calling requires the Responses API. Longer prompts, caching and processing tiers have their own billing conditions.
The useful enterprise question is not whether Terra replaces every larger model. It is whether a defined portion of a workflow can meet the same acceptance threshold with less total cost. Start with work whose outputs can be checked: interpreting a supplied document, preparing a structured draft, or proposing an application action for human approval. Keep the existing system as the control rather than changing model, instructions, tools and review policy simultaneously.
Independent evidence also needs a date and a metric. Artificial Analysis's current Terra/Mistral comparison reports Terra at Max effort with Intelligence Index v5.0 score 51, weighted cost of $0.96 per evaluated task and approximately 60,000 output tokens per task. These belong to that evaluation, not a price quote for invoice processing, a coding success rate, or a cost per accepted client deliverable.
Aeon's recommended pilot is to record acceptance and review time alongside the bill. An apparently cheaper first attempt can become expensive if it creates another review round. Conversely, a model that is not the highest-scoring option may be the right bounded route when its work is consistently acceptable. Neither outcome should be assumed from the published token rate alone.
2. Gemini 4 Carbon: separate capability from access
Google announced Carbon on October 28. Initial rollout is to selected cyber defenders and then trusted Fairwind partners; broad developer, business and consumer availability is still forthcoming. The announcement distinguishes introductory input/output rates of $2/$10 per million tokens from subsequent standard rates of $4/$20. We did not establish an end date for Carbon's introductory offer.
Artificial Analysis's October 28 evaluation article reports High-effort Intelligence Index v5 score 59, weighted task cost $7.76 and about 123,000 output tokens per task. Those costs use standard prices, not the introductory discount. Its separately reported coding-agent result is a different metric; identical-looking numbers do not make their task populations interchangeable.
For a procurement team, restricted access is a first-class constraint. A model cannot be an immediate replacement just because an evaluation makes it attractive. Confirm actual eligibility, the available endpoint, contractual data terms, service expectations and an accountable operational owner before including it in a production migration plan. A wait-list or announced rollout is not a production dependency you already control.
The bounded next step is an evaluation plan with an access gate. Specify the work, acceptance threshold and permitted data now, but retain the available production route until the new route is approved and tested. Cybersecurity capability claims are not evidence that autonomous tool access is safe in your environment. Authorization, logging, human escalation and rollback remain deployment decisions.
3. Mistral Large 4: downloadable weights change responsibility
Mistral's October 29 release makes improved Large 4 weights available after its earlier preview. The model card lists a one-million-token context window, approximately 1.05 trillion total parameters and 52 billion active parameters. Hosted API rates are $0.70 input, $0.07 cached input and $2.10 output per million tokens.
Open weights create a real option to control where and how inference runs. They do not establish an economical deployment on your current hardware. Capacity, utilization, serving software, resilience, patching and the team's operating time all belong in the comparison. Do not compare an API token price with a self-hosted server purchase and call the difference savings.
The published licence also matters. It is a modified MIT-style Mistral Ecosystem Open-Weight licence with conditions for model-as-a-service businesses and stated exceptions. Downloadable is not synonymous with unrestricted commercial redistribution. A proposed hosting or resale arrangement needs review against the actual text, not a benchmark site's shorthand openness label.
For the final High-effort model, the independent comparison above reports Intelligence Index v5.0 score 26, $2.06 weighted task cost and approximately 82,000 output tokens per task. Mistral's separate coding claims must not be paired with that Intelligence Index cost to manufacture a better operating point. The practical opportunity is a controlled hosting feasibility study, followed by a same-task pilot, not an assumed replacement for every frontier API.
A buyer's decision summary
| Decision | First gate | Evidence to collect |
|---|---|---|
| Route work to Terra | Same output-acceptance threshold | Total billed usage, retries and reviewer minutes |
| Evaluate Carbon | Verified access and data terms | Same-task quality, availability and durable pricing |
| Host Large 4 | Licence and operating feasibility | Utilization, infrastructure, support and accepted work |
The Lab's default now contains 18 models. Our same-provider rotation moves GPT-6 Astra and Gemini 3.8 Flash out of the default only, based on verified release chronology. Both remain in All / history. Mistral introduces a provider family with no existing default to replace. All historical operating points are retained.
The three additions are explicitly provisional, cross-suite Intelligence Index diamonds. They are not directly rank-comparable with the historical coding-suite circles, and no global cross-suite winner is inferred. Current API rates, benchmark task costs and production cost per accepted work remain separate quantities.
Before changing a supplier, define the business acceptance test and keep a record of what passed. Aeon's AI Control and ROI Assessment can help turn model choice into a bounded workflow, cost and control decision. Contact info@airiskmanagement.ca to discuss the scope. No benchmark in this article establishes your organization's savings or control effectiveness.