--- title: "69 Releases in a Week, and Why Model Switching Compounds" slug: "sixty-nine-releases-model-agnostic-platform-compounds" author: "Blanca Amigot" date: "2026-09-10 15:00:00" category: "Premium" topics: "model-agnostic, LLM registry, model deprecation, self-hosted AI, OpenAI-compatible API, enterprise AI, shipping velocity" summary: "ibl.ai shipped 69 web frontend releases in the week to September 4, 2026, refreshing its LLM registry to GPT-5.6, Claude Opus 5, Gemini 3.7 and DeepSeek V4. Models retire on the provider's calendar, not yours." banner: "" thumbnail: "" linkedin: | Sixty-nine web frontend releases in the week to September 4, 2026. The most useful line in that changelog is not a feature. It is the LLM registry refresh. Every model in a production system has a retirement date set by someone else. OpenAI shuts down gpt-4.1-nano, gpt-4-turbo, o1, o3-mini and o4-mini on October 23, 2026. Anthropic commits to at least 60 days' notice and used close to the minimum on Claude Opus 4.1: deprecated June 5, 2026, retired August 5. Google retired the stable Gemini 2.0 Flash models on June 1, 2026. So the honest framing is not that models change quarterly. It is that the model you shipped on gets switched off on a date you did not choose, sometimes with two months' warning. That should be a configuration change, not a project. → An LLM registry seeded from one command, so adding GPT-5.6, Claude Opus 5, Gemini 3.7, Grok 4.6 or DeepSeek V4 is a data operation, not a code change → Model choice per agent, plus a default that tracks the best available model rather than pinning a two-year-old one → A standards-compliant OpenAI-compatible /v1 endpoint, so application code keeps talking to the same SDK while the model underneath changes → An on-device provider option, for workloads that should never leave the machine With ibl.ai you own all the code and the data — self-hosted inside your own perimeter, model-agnostic across any LLM, usage-based with no per-seat pricing, deployable anywhere from your own cloud to a fully air-gapped network. #iblai #AgenticAI #EnterpriseAI #LLMOps #ModelAgnostic #SelfHosted --- ## The Short Answer **ibl.ai shipped 69 web frontend releases between August 28 and September 4, 2026, refreshing its LLM registry to GPT-5.6, Claude Fable 5, Opus 5 and Opus 4.8, Gemini 3.7, Grok 4.6 and DeepSeek V4. Models turn over on the provider's schedule: OpenAI retires gpt-4.1-nano on October 23, 2026. With ibl.ai you own all the code and the data, so switching models is a registry entry.** The models in that list will all be superseded. The mechanism for replacing them is the part that keeps paying. ## What actually shipped in ibl.ai's September 4 platform update? Sixty-nine releases across the web frontend, covering **2026-08-28 through 2026-09-04**, over platform pins `1.174.0` to `1.198.0`. The full changelog is published at [the September 4 platform update](/updates/platform-update-2026-09-04). One correction to how this gets repeated. That is a count for one week from one changelog, not a sustained rate, and it is not a benchmark against anyone else's release process. Release counts are only interesting for what they let you do. The entries that matter for model turnover are four. The **LLM registry** was refreshed to current provider lineups — OpenAI GPT-5.6, Anthropic Claude Fable 5 / Opus 5 / Opus 4.8, Google Gemini 3.7 and 3.6 Flash tiers, xAI Grok 4.6 and 4.5, DeepSeek V4, with matching Azure and Bedrock entries. Two rows were removed because their providers deprecated them: Perplexity's `sonar-reasoning` and DeepSeek's `deepseek-chat`. A new default model, **ibl.ai Pro**, replaced `gpt-4o-mini` for newly created agents, so a new agent tracks the best available model instead of pinning a two-year-old one. A standards-compliant **OpenAI-compatible `/v1` endpoint** shipped with Bearer-key auth, a structured error envelope and tenant-scoped model listing. And the LLM tab gained an **on-device provider** picker. ## How fast do frontier models actually turn over? Faster than "quarterly," and the number that binds is not the release cadence. It is the retirement date. OpenAI's deprecations page schedules a block of legacy snapshots — `gpt-4.1-nano`, `gpt-4-turbo`, `gpt-3.5-turbo-0125`, `o1`, `o3-mini`, `o4-mini` and their fine-tuned variants — for [shutdown on October 23, 2026, announced April 22, 2026](https://developers.openai.com/api/docs/deprecations). Anthropic [commits to at least 60 days' notice](https://platform.claude.com/docs/en/docs/about-claude/model-deprecations) before retiring a publicly released model, and has run close to that floor: Claude Opus 4.1 was deprecated on June 5, 2026 and retired on August 5, 2026. Google [retired the stable Gemini 2.0 Flash and Flash-Lite models on June 1, 2026](https://ai.google.dev/gemini-api/docs/deprecations), pointing Flash users at `gemini-3.6-flash` and Flash-Lite users at `gemini-3.1-flash-lite`.
Provider Example retirement Announced Switched off
OpenAI gpt-4.1-nano, gpt-4-turbo, o1, o3-mini, o4-mini Apr 22, 2026 Oct 23, 2026
Anthropic claude-opus-4-1 (60-day minimum notice policy) Jun 5, 2026 Aug 5, 2026
Google gemini-2.0-flash, gemini-2.0-flash-lite not stated Jun 1, 2026
Read that as an operations calendar rather than a news cycle. A model you shipped on gets switched off on a date another company chose, and in the tightest case you get about two months to move. ## What does it take to switch models on a platform you run yourself? Three things, and none of them should be an application rewrite. **A registry, not a code path.** In the September 4 release, `seed_llm_registry` became the single command that populates providers and models, replacing per-flow JSON files. Adding a provider or retiring a dead row is a data operation you run against your own deployment, with a `--prune` pass to move agents off models that no longer exist. **Model choice per agent.** Different work wants different models: a long-context research agent and a high-volume classifier have no reason to share one. Selection sits on the agent, so changing one agent's model does not touch the others. **A default that moves.** The new-agent default is now ibl.ai Pro, which tracks the best available model. The failure mode it removes is the pinned default nobody revisits, which is exactly how a fleet ends up on a model with a shutdown date. The same argument runs through [GPT-6 Astra's benchmark harness footnote](/blog/gpt-6-astra-arc-agi-3-model-agnostic-architecture): the score that matters depends on the scaffolding around the model, so the scaffolding is the thing worth owning. ## Why does an OpenAI-compatible endpoint matter for switching models? Because it keeps the switch out of your application code. The `/v1` endpoint that shipped on September 4 is standards-compliant: Bearer-key auth, a structured error envelope, and a tenant-scoped model listing. Any OpenAI SDK can be pointed at your own deployment. That inverts the usual dependency. The SDK your teams already use becomes a stable interface to your platform, and the model behind it is a setting. You are not migrating client libraries every time a provider sunsets a snapshot. The on-device provider option covers the other end of the range, for workloads that should not leave the machine at all — which is a model-selection question with a data-residency answer. ## Why is shipping velocity different when you host the platform yourself? Because you take releases on your own maintenance window rather than on the vendor's. On a managed SaaS product, an upstream change arrives when the vendor deploys it. Your change-control process is informed, not consulted. On a self-hosted platform the same 69 releases arrive as a pin you choose to move to, after you have tested it, in a window your operations team scheduled. That is why the September 4 update lists migrations, index builds and a `--prune` step by name. Those are instructions for an operator who controls the deploy, not release notes for a user who receives one. ## How does ibl.ai make model switching something you own? With ibl.ai you own all the code and the data. The platform runs on your own infrastructure with full source code access, is model-agnostic across any LLM so you can add or drop a provider in the registry without rewriting the platform, is usage-based with no per-seat pricing, and can deploy anywhere — your own cloud, on-premise, GovCloud, or a fully air-gapped network with no outbound connectivity. The practical consequence is that a provider's deprecation notice becomes a scheduled change instead of a dependency. 1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University. ibl.ai is family-owned and operated from New York, NY. *Related reading: [GPT-6 Astra, ARC-AGI-3, and the harness footnote](/blog/gpt-6-astra-arc-agi-3-model-agnostic-architecture) — why the architecture around a model outlasts any single benchmark result.* *Sources: release counts, the LLM registry refresh, the default-model change and the OpenAI-compatible `/v1` endpoint from ibl.ai's own [September 4, 2026 platform update](/updates/platform-update-2026-09-04); the October 23, 2026 shutdown of `gpt-4.1-nano` and other legacy snapshots from [OpenAI's deprecations page](https://developers.openai.com/api/docs/deprecations); the 60-day notice policy and the Claude Opus 4.1 dates from [Anthropic's model deprecations page](https://platform.claude.com/docs/en/docs/about-claude/model-deprecations); the June 1, 2026 Gemini 2.0 retirement from [Google's Gemini API deprecations page](https://ai.google.dev/gemini-api/docs/deprecations).* ## Why does owning the AI stack matter? **ibl.ai is the agentic AI platform where you own all the code and the data. You self-host the entire stack inside your own perimeter, run it model-agnostic across any LLM and switch anytime, and pay by usage with no per-seat pricing — so you can deploy anywhere: your cloud, on-premise, GovCloud, or fully air-gapped.** - **You own all the code and the data.** Full source code under a perpetual license, running on your infrastructure. Not API access to someone else's platform — the stack itself is yours. - **Model-agnostic.** Run any LLM — Claude, GPT, Gemini, Llama, Command, or your own fine-tune — and switch providers without rewriting the platform. - **No per-seat pricing.** Usage-based billing against a budget cap you set. Cost tracks what your organization actually uses, not how many people you employ. - **Deploy anywhere.** Your cloud, your VPC, on-premise, GovCloud, or a fully air-gapped network with no outbound connectivity. 1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University. ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on.