Deprecation Notice #001: GPT-5.6 Ships, Gemini Slips, and the Job Gets Renamed Again
Par Salty Deprecated Software Engineer
Cet article a été généré avec l'aide de l'IA et revu par notre équipe pour en garantir l'exactitude et la qualité. Toutes les informations techniques et exemples ont été vérifiés.
Welcome to the first Deprecation Notice: the weekly digest of what shipped in AI, what it breaks, and what someone who actually operates systems should do about it. No countdown of "mind-blowing tools." Just the changelog, read the way you'd read a vendor's — skeptically, with the invoice in the other hand.
Format note: every item carries a dated source. Anything we couldn't verify got cut — including one story you may have seen elsewhere this week. Details at the bottom.
OpenAI Ships GPT-5.6 — Now With Three SKUs, Like a Server Line
On July 9, OpenAI launched the GPT-5.6 family in three sizes: Luna ($1 input / $6 output per million tokens), Terra ($2.50 / $15), and Sol ($5 / $30). If that product ladder looks familiar, it should — it's the same good/better/best play every hardware vendor has run since the ProLiant era, except the line items are tokens now instead of drive bays. (TechCrunch, Jul 9; Axios, Jul 9)
Re-run your monthly numbers before anyone on the team "just switches to Sol." A 5x input-price spread across one model family is a capacity-planning problem, and capacity planning is your job description with new nouns. Our AI Cost Planner takes requests/day and tokens/request and gives you the bill per model — the pricing table in it carries a last-verified date, and this release is exactly why.
Meta Starts Charging — Muse Spark 1.1 and a Paid API
Also July 9: Meta released Muse Spark 1.1 and, for the first time, a paid developer API — public preview, US-only, $1.25 input / $4.25 output per million tokens, $20 in free credits, 1M-token context window. The company that spent a decade giving models away just joined the metered-billing business. Free tiers are a customer-acquisition phase, not a personality trait. (Meta AI blog, Jul 9; MarkTechPost, Jul 9)
Treat it like any other vendor eval: burn the $20 of credits on your own workload, not their demo. A 1M-token window sounds infinite until you check what your actual documents cost to stuff into it — paste one into the Token Counter and see how much context you're really buying per call.
Kimi K3: "Open Weights" Now Means 2.8 Trillion Parameters
Moonshot AI released Kimi K3 on July 16 — a 2.8-trillion-parameter mixture-of-experts model (16 of 896 experts active per token) that took the #1 spot on Arena.AI's Frontend Code Arena ahead of the closed frontier models, with full weights promised by July 27. It is, by parameter count, the largest open-weight model ever shipped. (Tom's Hardware, Jul 16; VentureBeat, Jul 16)
Reality check: "open weights" and "runs on your hardware" stopped being the same sentence a while ago. 2.8T parameters is a rack budget, not a workstation download. Downloading the weights makes you a mirror, not an operator.
If the K3 headlines have someone asking "can we self-host this?", answer with math instead of a meeting. The Local AI Checker shows what your RAM/VRAM actually runs, 3B to 70B, with the memory arithmetic visible — which is the honest end of the local-model conversation.
Google Delays Gemini 3.5 Pro Over Coding Performance
Bloomberg reported July 16 that Google has pushed back the broader release of Gemini 3.5 Pro after internal testing fell short on coding and long-horizon reasoning — a late-June training-data update aimed at fixing it reportedly disappointed. The model previewed at I/O stays in limited enterprise preview while an upgraded Flash model is tested with partners. (Bloomberg, Jul 16; 9to5Google, Jul 16)
If any plan on your desk says "when 3.5 Pro lands," add a date column and a fallback. You wouldn't schedule a migration around unreleased firmware; a previewed model is unreleased firmware with a keynote. Anything you operate that states model capabilities should carry a verified date — ours do, and this is the week that discipline earned its keep.
The Item We Cut (And Why That's the Whole Point)
An aggregator making the rounds this week claimed Anthropic had just yanked Claude Code from its $20 Pro plan and reversed course after backlash. Compelling story. Wrong dateline: that A/B test and walk-back happened in April 2026, and the separate agent-billing change was paused back in June. (BigGo, Apr 2026; Zed blog)
So it got cut. That's the deal with this digest: if a claim can't be pinned to a dated source, it doesn't run — the same rule in our editorial policy that governs everything else on this site. Check the dateline before you forward the outrage. Old ops rule, new failure mode.
Until Next Week
That's issue #001. The pattern going forward: what shipped, what it costs, what it breaks, what you do Monday morning. If you spot an error, the corrections policy tells you exactly how to make us fix it — reported outages are how systems get better, and the reporter is never the problem.
Écrit sous le nom de plume éditorial de The IT Hustle : plus de 25 ans comme technicien laptop, administrateur système, ingénieur stockage et ingénieur logiciel, aujourd’hui aux commandes d’agents IA. Chaque article est relu par un humain avant publication ; voir la charte éditoriale.
Reste dans la boucle
Sois le premier informé des nouveaux outils, articles et mises à jour. Sans spam.
Générez vos propres prompts anti-hallucination
Notre AI Prompt Engine utilise une technologie propriétaire pour générer des prompts avec vérification intégrée et tests de contradiction.
Essayer 3 générations gratuites →