This week's model-buying decisions are about what you can use, what you must migrate, and what your agent actually bills. Thinking Machines has introduced Inkling 2 Small through Tinker, with downloadable weights still conditional. Google has set a replacement-routing deadline for Gemini 2.5 Pro and Flash. DeepSeek's previously announced Search and Fetch billing has taken effect.
Our November 6 review covers October 31 through November 6, without extending the seven-day window. Google's notice is dated November 2; DeepSeek's relevant change is the October 31 billing boundary, not a new model launch. These are sourced developments and Aeon's analysis, not measured Aeon client results. The Model Economics Lab separates published evaluation observations from the operational decisions below.
1. Inkling 2 Small: available through Tinker, not open weights today
Thinking Machines announced Inkling 2 Small on November 5. Its model card distinguishes current closed-weight access through Tinker from an intended Apache 2.0 weight release the following week, conditional on external safety testing. A release intention is not a downloadable artifact or a licence you can exercise today.
Tinker's models and pricing documentation lists a 256K context and serverless inference rates of USD $0.30 input, $0.03 cached input and $1 output per million tokens. Serverless inference is in beta with a waitlist. A buyer should therefore check actual access before treating those rates as a deployable alternative. The context limit also does not establish reliable retrieval or answer quality across a full business document set.
Artificial Analysis's Inkling evaluation reports Intelligence Index v5.0.1 score 31, weighted cost of $1.18 per evaluated task and approximately 150,000 output tokens per task. We did not verify the evaluated thinking-effort setting. Those are independent evaluation observations, not 31% coding accuracy, a token rate, or a predicted bill for your workflow.
The Lab adds a provisional cross-suite observation for Inkling rather than presenting a new same-harness coding measurement. Thinking Machines is a new provider in this dataset, so the default grows from 18 to 19 models without evicting another provider. Historical points remain unchanged. Its marker is an invitation to inspect evidence, not a recommended routing setting or proof of savings.
For teams already able to access Tinker, the bounded action is a small shadow evaluation against their current model on non-sensitive, checkable tasks. Record access constraints, task acceptance, reviewer minutes, retries and billed usage. Do not design a private deployment around next week's conditional weight release. Revisit that decision only when the actual artifacts, terms and operational requirements can be inspected.
2. Gemini 2.5 migration: choose your target before automatic routing
Google's November 2 Gemini API notice deprecates Gemini 2.5 Pro and Gemini 2.5 Flash and removes them from AI Studio selection. Effective December 1, API requests to `gemini-2.5-pro` will automatically route to Gemini 3.8 Flash; requests to `gemini-2.5-flash` will route to Gemini 3.5 Flash-Lite. Replacement-model rates will apply. This is an existing-model migration notice, not a new model launch.
The notice identifies breaking API changes involving function calling and `thought_signature` validation, `thinking_level` replacing `thinking_budget`, and removed sampling parameters and prefilled model turns. An unchanged request URL or an automatic redirect is therefore not evidence of a like-for-like workflow. A response that still arrives can differ in quality, tool behavior, latency and bill.
Start with an inventory of model strings, aliases and defaults in application configuration, scheduled jobs, SDK wrappers and external tools. Assign an owner to each affected route. Make an explicit target choice against the workflow's acceptance criteria instead of letting the calendar make the selection. The published replacement is a migration path to test, not a recommendation for every business task.
Use a bounded regression set that includes ordinary answers, structured outputs, multi-turn tool calls, error handling and representative long prompts. Check that function arguments remain valid, required signatures survive the conversation, and reasoning controls map to the new API rather than being silently dropped. Compare accepted results, reviewer time, retries, latency and actual usage under the replacement rate card. Do not change prompts, model and review policy together unless the migration requires it; record necessary changes separately.
Plan the cutover before December 1, with a tested fallback and an observable failure path. Neither the notice nor a lower-looking token rate establishes savings or comparable capability. This migration and DeepSeek's billing change create no new Lab point or default-model retirement, and existing historical observations remain unchanged.
3. DeepSeek Search and Fetch: the free period has ended
DeepSeek's October 17 announcement set October 31, 2026 at 16:00 UTC as the start of billing for previously free server-side Search and Fetch tool calls. This week's event is that effective date, not a newly launched model or newly announced search service.
The Search and Fetch guide lists each endpoint at USD $0.75 per 1,000 successful requests; failed or empty requests are unbilled. Search and Fetch are separate billable endpoints. Model-token charges remain additional. A research agent that searches and then fetches several pages therefore needs both counters, not a single generic “web search” allowance.
The buyer implication is accounting discipline. A low token bill can hide tool usage, while aggressive limits can impair evidence quality. Record successful searches, successful fetches, model usage, retries, accepted answers and review time separately. Capture which requests actually produced useful evidence, rather than rewarding an agent for collecting more pages.
The bounded pilot is to instrument a small existing research workflow, set request limits and compare its cost per accepted answer before changing providers. Preserve source-quality and freshness checks. This update supports a budget adjustment; it does not by itself establish that DeepSeek is cheaper than another retrieval stack or that paid calls improve answer quality.
Which decision belongs on your desk?
| Change | Immediate buyer decision | Evidence still needed |
|---|---|---|
| Inkling 2 Small | Evaluate hosted access if available | Task acceptance, actual access, and any eventual weight-release terms |
| Gemini 2.5 replacement routing | Inventory affected routes and choose a tested target before December 1 | API compatibility, task acceptance and replacement-rate billing |
| DeepSeek Search and Fetch billing | Update agent budgets and usage telemetry | Tool counts and cost per accepted, adequately sourced answer |
Google's deadline makes affected-system inventory the most time-sensitive action. Inkling access and DeepSeek usage accounting can be evaluated without replacing a working system this week. Keep each comparison narrow: define an acceptable answer before testing, hold inputs and review rules constant, and record all attempts rather than only successful examples. Use approved data and explicit spending limits; this research issue did not run paid API or benchmark jobs.
The shared measure is total cost per accepted task: inference, tools, retries, infrastructure where relevant, and human review. Keep quality and service-level thresholds alongside it so a cheaper but unusable result cannot win. Published benchmark costs help identify candidates; they do not supply that denominator for your business.
Use the Lab's evidence and methodology to inspect the provisional Inkling observation and unchanged history. For help defining a migration acceptance test or a bounded model-cost assessment, contact Aeon. The useful next step is an owned decision with evidence, not an assumed saving from a launch headline.