Market notes — The Marginal Tokenhttps://marginaltoken.com/Changes that affect the cost of buying and running AI.enFri, 23 Oct 2026 02:01:29 GMTFlint: local coding, hardware requirements and running costshttps://marginaltoken.com/notes/flint-local-coding-hardware-running-costs/urn:marginal-token:note:flint-local-coding-hardware-running-costsThu, 22 Oct 2026 00:00:00 GMTRecursive’s coding model has downloadable Apache-2.0 weights. Its optimized Mac runtime requires specific chips and at least 48 GB of memory; running costs depend on hardware use and task success.Clairic: extraction at $0.10 per million input tokenshttps://marginaltoken.com/notes/clairic-extraction-input-only-pricing/urn:marginal-token:note:clairic-extraction-input-only-pricingWed, 21 Oct 2026 00:00:00 GMTCloudflare’s extraction preview charges for input only. The practical test is cost per accepted extraction, including validation and human review.Mercury 3: launch prices and the Mercury 2 migration deadlinehttps://marginaltoken.com/notes/mercury-3-launch-pricing-migration/urn:marginal-token:note:mercury-3-launch-pricing-migrationTue, 20 Oct 2026 00:00:00 GMTMercury 3 launches at half its regular token rates. Existing direct customers of Mercury 2 and 2.5 have until 15 November to migrate, with enterprise agreements treated separately.AI model releases: 12–16 October 2026https://marginaltoken.com/notes/ai-model-releases-2026-10-12/urn:marginal-token:note:ai-model-releases-2026-10-12Fri, 16 Oct 2026 00:00:00 GMTGemini 4 Argon and Muse Wonder expand the API choices. Muse Spark Open and NVIDIA’s Nemotron releases bring new deployment options. A review of the week so far, with access and costs checked.Muse Spark Open: deployment hardware and license termshttps://marginaltoken.com/notes/muse-spark-open-rider/urn:marginal-token:note:muse-spark-open-riderFri, 16 Oct 2026 00:00:00 GMTMuse Spark Open’s weights are available under custom terms. Compare its deployment hardware and licensing conditions with hosted API prices.Nemotron 3.5 Super VL and Bolt: model size and local deploymenthttps://marginaltoken.com/notes/nemotron-super-vl-bolt-packs-light/urn:marginal-token:note:nemotron-super-vl-bolt-packs-lightFri, 16 Oct 2026 00:00:00 GMTNVIDIA’s Bolt compresses Super VL for local deployment. A smaller model footprint changes the hardware calculation; it does not settle the bill.Step 5 Preview: weights, hardware and license termshttps://marginaltoken.com/notes/step-5-preview-entourage/urn:marginal-token:note:step-5-preview-entourageThu, 15 Oct 2026 00:00:00 GMTStep 5 Preview’s weights are public. Its active parameter count tells only part of the deployment story.Gemini 4 Argon: API prices double on 1 Decemberhttps://marginaltoken.com/notes/argon-introductory-rate/urn:marginal-token:note:argon-introductory-rateWed, 14 Oct 2026 00:00:00 GMTGemini 4 Argon’s posted token prices double on 1 December. Budget recurring work using both the introductory and scheduled rates.Muse Wonder: API pricing and Wonder Full accesshttps://marginaltoken.com/notes/muse-wonder-booking-terms/urn:marginal-token:note:muse-wonder-booking-termsTue, 13 Oct 2026 00:00:00 GMTMuse Wonder’s API rates match Opus 5.5. Wonder Full is available through specific Muse Code subscriptions.Haiku 5.5: higher API rates above 100,000 prompt tokenshttps://marginaltoken.com/notes/haiku-5-5-context-pricing/urn:marginal-token:note:haiku-5-5-context-pricingMon, 12 Oct 2026 00:00:00 GMTThe 100,000-token boundary changes both input and output rates. Price agent workloads by prompt length, with cached tokens included.