Local AI and Token Pricing Take Center Stage

ByMason Reed

November 22, 2026

Thinking Machines’ open-weight Inkling 2 Mini targets local coding work, while Google’s Gemini 4 Neon offers developers a clearly priced, long-context API option.

The clearest developer-facing AI news in the latest available material is practical: Thinking Machines has released an open-weight model designed for local execution, while Google has published detailed capabilities and pricing for its Gemini 4 Neon API. Both give software teams more options to weigh against cloud dependence, operating costs and hardware limits.

Thinking Machines released Inkling 2 Mini on November 20; VentureBeat covered it the next day. The company describes it as a compact, multimodal model intended to run locally while handling coding and agentic tasks. Its model card lists 26.4 billion total parameters, 3.4 billion active parameters, an approximately 20GB memory footprint and a one-million-token context window. That combination makes it a substantial local model, not a lightweight tool that will run comfortably on every laptop.

Inkling’s looped-transformer design runs its 12 middle layers three times for each token. That adds internal computation without requiring proportionally longer generated reasoning. Developers can run it through Transformers, llama.cpp and MLX, with SGLang support for production GPU inference. These integrations matter because a model’s usefulness depends not only on its weights but also on whether it fits existing developer workflows. The open-weight release and runtime options make it a candidate for teams that want to experiment without sending every prompt to a cloud provider. Local deployment can also give businesses more control over where code and other data are processed.

The model’s published coding results include 48.2% on DeepSWE v1.1, 77.9% on SWE-bench Pro v2 Hard, 79.2% on SWE-bench Verified and 46.5% on Vals Index v2.2. These figures come from company or model-card materials reported by VentureBeat and The Next Web; they are not independently verified major-outlet benchmarks. They are a reason to test the model, not a guarantee it will fix bugs or complete tasks reliably in a particular codebase. Teams should also assess speed and hardware needs: a stated memory footprint does not by itself establish practical performance, and local inference still has a power and setup cost.

Google’s Gemini 4 Neon is another concrete option. Google AI Studio documentation lists four thinking levels—low, medium, high and xhigh—and a one-million-token context window, with decoding of up to one million tokens. Google’s paid API pricing is $3 per million input tokens and $15 per million output tokens. Cached input is listed at $0.15 per million. The specialist outlet The Inference reported general availability on November 18 at the same rates.

Published pricing helps developers estimate costs before building around a model. A large context window may suit projects involving extensive code or documents, but it does not settle questions of latency, accuracy or token consumption. The low cached-input rate may be relevant for repeated prompts or workflows that reuse context, though actual savings depend on how an application is built. Teams considering Neon should compare it with existing providers on representative tasks and actual bills, rather than relying on headline benchmarks. VentureBeat also describes Google’s lineup as spanning cost and performance tiers, but the available pricing details here are specifically for Neon.

The available material offers less certainty about a broader claim that OpenAI has just changed the coding-model market. It does not verify a major-outlet report from November 21–22 confirming a GPT-6.2 Astra launch. A specialist outlet’s reported price comparisons are not enough to establish a like-for-like market picture. The supplied reporting also falls short of verifying three to five major-outlet AI stories from the last 24 hours, so a fuller ranking of the day’s launches would overstate what is known.

For builders, the practical next step is evaluation, not wholesale adoption. Inkling 2 Mini is worth trying for teams interested in open weights and local inference; Gemini 4 Neon merits a cost and performance comparison for teams seeking a hosted model with documented pricing and long context. Test both against real tasks, data-handling requirements and budgets before making them part of a software stack.

Leave a Reply

Your email address will not be published. Required fields are marked *