AI Developer News: Week Ending October 17, 2026

This week’s AI developer news is a story of endings and openings at the same time: OpenAI is pulling GPT-4o and GPT-4.1 out of its API, GitHub Copilot is retiring three more models in November, and — in the same breath — four open-weight models landed from NVIDIA, Reflection AI and Liquid AI, including a 501B-parameter coding model under a full Apache 2.0 licence. Anthropic spent the week having Claude design quantum circuits that beat published state-of-the-art, and ChatGPT Voice got a new default voice model with 13 new voices. Here’s what each one actually means for the code you ship — and what to do about it this week.

This AI developer news roundup covers the week ending October 17, 2026. Every item is dated and sourced from the original announcement or the strongest independent coverage, and unverified vendor claims are labelled as such.

Table of Contents

OpenAI is killing GPT-4o and GPT-4.1 in the API

The deadline you should act on first. OpenAI added a new entry to its API deprecations page on October 14, 2026: five model snapshots leave the API on January 14, 2027 — gpt-4o-2024-08-06, gpt-4o-2024-11-20, gpt-4o-mini-2024-07-18, gpt-4.1-2025-04-14 and gpt-4.1-mini-2025-04-14. Fine-tuned versions of these models shut down on the same date — there is no exemption for fine-tunes. OpenAI’s recommended replacement for all of them is gpt-6-luna.

The timetable is actually tighter than the headline suggests. Two earlier snapshots — gpt-4o-2024-05-13 and gpt-4.1-nano — leave on October 23, 2026, less than a week from now. The TTS models (tts-1, tts-1-hd, gpt-4o-mini-tts) go on January 6, 2027, and the GPT-4o audio and realtime models follow on January 20, 2027.

One caveat worth knowing: this is only about three months’ notice, while OpenAI’s own policy normally promises six months for generally available models. The company cites a safety/compliance exception but tied it to no specific reason. Whatever the justification, the date is fixed. If any production workflow, classifier, extractor or agent still calls a gpt-4* model string, you have roughly three months to migrate — and for the nano and May-2024 snapshot users, roughly one week. Grep your configs this weekend.

GitHub Copilot retires three more models on November 16

A second deprecation wave, this time from GitHub. The GitHub Changelog announced on October 15 that three more models retire on November 16, 2026 across all Copilot experiences — Chat, inline edits, ask and agent modes, and code completions:

Claude Opus 4.8 → Claude Opus 5.5 (exception: Opus 4.8 stays available for individual Copilot subscribers on annual plans, so annual subscribers keep an Opus option)
Claude Opus 5 → Claude Opus 5.5
GPT-5.6 Terra → GPT-6.1 Sol or GPT-6 Luna

This is a genuinely new wave, not a re-announcement: it follows the October 2 wave (Opus 4.7, Gemini 3.5/3.6 Flash, Kimi K2.7 Code) and is separate from the six-model October 19 retirement covered in last week’s roundup. Under default model enablement, the replacements are auto-enabled for Copilot Enterprise and Business seats unless an admin disabled the global default. But the deprecation hits inline edits and completions too, not just the Chat picker — so audit pinned model strings in IDE defaults, agent-mode configurations, eval harnesses and Copilot CLI scripts, not just the dropdown. If your team runs Copilot’s local/hybrid setups, review which models are actually serving your requests while you’re at it.

NVIDIA ships Nemotron 3.5 Super VL and Bolt as open models

NVIDIA joined the open-weight week on October 15 with three agent-focused open models under its own permissive licence:

Nemotron 3.5 Super VL — 120B total / 12B active MoE, multimodal (text, images, short video, PDFs), 1M-token context, built for enterprise multimodal agents. Available now.
Nemotron 3.5 Bolt 74B-A9B — 74B total / 9B active MoE with native 2-bit quantization, so it fits in a single 24GB card — RTX 5090-class PCs, RTX and DGX Spark hardware. 256K context. Available now.
Bolt E4B 4GB — an edge/conversational agent model due October 23.

Weights are on Hugging Face alongside training datasets and recipes, with day-one support in llama.cpp, vLLM, Ollama, Unsloth and LM Studio. The licence is NVIDIA’s own OpenMDW-1.1 — permissive with no revenue gates, but it’s NVIDIA’s licence, not an OSI-approved one, so have legal glance at it before commercial use. NVIDIA claims Super VL matches its largest reasoning model at roughly one-fifth the parameters with 5x throughput; Artificial Analysis scores Super VL at 26 on its Intelligence Index — treat both as vendor and third-party snapshots rather than settled truth. The genuinely newsworthy bit for builders: a 74B-class MoE agent model that fits one consumer GPU is the first practical “local agent brain” from a first-party vendor. Prototypes that needed cloud inference can now stay on-device.

Reflection AI open-sources Beam 501B weights under Apache 2.0 — plus the Mirror CLI

The most permissive licence of the open-weight frontier race so far. Reflection AI released the weights of Beam — 501B total / 23B active MoE, text-only, 1M-token context, built for coding, reasoning and agentic work — under Apache 2.0 on October 16 (the model itself was announced on October 5). Full commercial use, no revenue gates. That makes it the lawyer-friendly, fine-tunable alternative for teams that can’t or won’t use Chinese-lab models.

The release package is serious: BF16, FP8 and NVFP4 formats, a companion DFlash2 drafter model for speculative decoding, a first-party API beta at platform.reflection.ai with an OpenAI-compatible client, and day-one hosted support on Baseten, Fireworks, Together, OpenRouter, Nebius, Modal and NVIDIA NIMs. Reflection also open-sourced Mirror CLI, the terminal harness it used to build and train Beam — it runs parallel agents in separate Git worktrees, which will look familiar if you read our git worktrees for parallel AI agents guide.

Two caveats before you plan a migration. First, the vendor benchmarks (80.9% on SWE-bench Verified) are unconfirmed by independent reproduction — Artificial Analysis’s third-party read is more modest: 30 on its Intelligence Index, the #2 US open-weight model behind Meta’s Muse Spark Open at 39, and notably token-efficient at ~26k output tokens per index task. Second, Beam is data-centre-class hardware (~500GB at 8-bit), not a laptop model. If you want open weights on a consumer GPU, NVIDIA’s Bolt is your target; if you want the most capable permissive-licensed open model, Beam is now the answer.

Liquid AI releases LFM2.5-24B-A2B for on-device agents

The efficiency play of the week. Liquid AI released LFM2.5-24B-A2B on October 13: 24B total / 2.45B active per token MoE (64 experts, top-4 routing), reasoning-only — it always emits chain-of-thought — with a 131K context window (up from 32K) and a 128K-token vocabulary, trained on roughly 33T tokens. Both base and post-trained checkpoints are on Hugging Face.

What distinguishes it is the training method: post-training included agentic reinforcement learning inside real agent harnesses — OpenClaw, Hermes Agent, Pi and OpenCode — so it’s a tool-calling brain shaped by the environments it’ll actually run in. Liquid claims 184 output tokens per second on an Apple M5 Max. Vendor benchmarks against Qwen3.6-35B-A3B are split: Liquid wins on AutomationBench (27.7 vs 8.7) and LiveClaw (48.1 vs 38.2), loses on SWE-bench Verified (51.0 vs 66.2) and GPQA Diamond (74.4 vs 83.7) — read it as a strong agentic specialist rather than an all-rounder. Two warnings: the licence is Liquid’s own lfm1.0, not Apache/MIT, so check commercial terms; and there is a known Apple-Silicon quirk — older mlx-lm versions route its experts incorrectly, so update your stack before testing.

Anthropic: Claude Mythos 5 designs quantum circuits beating the published best

The week’s most interesting research item for agent builders. Anthropic published a research post on October 16 (with a companion 46-page arXiv paper posted October 15 by Andi Gu, Kemeny, Yelin of Harvard and Mikhail Lukin of Harvard/QuEra) showing Claude Mythos 5, running inside the open-source evolutionary search OpenEvolve, designing quantum circuits that beat the best published designs.

The loop is the story: Claude wrote each circuit as Python, an automated evaluator checked correctness and scored it by Toffoli-gate × qubit cost, with no human input between rounds — a closed propose → evaluate → refine pattern. Across eight primitives (chemistry, optimisation, physics simulation, number theory), 7 of 8 tasks reduced both Toffoli gates and qubit counts versus the best published designs, with up to 39x cost reduction on one task; the eighth yielded a new reversible architecture for prime-field modular inversion. The headline run: 53 hours on elliptic-curve point addition produced a circuit with 12% more qubits but 17x fewer Toffoli gates than the smallest previous design, and a Reed–Solomon decoder cut gate count 12x versus Google Quantum AI’s construction. Claims were checked via classical simulation and Lean formal proofs.

Why it belongs in a developer news roundup: this is the concrete template for AI-for-science and automated eval-driven agent loops — an agent plus a cheap, checkable grader producing publishable results beyond chat. Caveats apply: it’s a single preprint with no independent reproduction yet, the figures are resource-count estimates for hypothetical fault-tolerant machines rather than hardware runs, and no — this has no implication for elliptic-curve crypto safety; it’s one arithmetic component, still needing millions of gates on hardware that doesn’t exist.

ChatGPT Voice gets GPT-Live-1.1 and Quip, with 13 new voices

OpenAI rolled out GPT-Live-1.1 in ChatGPT on October 15, replacing GPT-Live-1 as the default voice model for Plus and higher tiers (including Codex and Dots), while GPT-Live-1.1 Quip (the mini) replaces GPT-Live-1 mini for other tiers. The update adds 13 new voices, new language options for French, German, Hindi, Japanese, Korean and Mexican Spanish (plus an English voice called “Bubbie”), more natural accents and more varied short phrases while the model searches. It’s full-duplex — it listens and responds continuously.

The limits that matter if you test it: Go plan gets 3 hours per 24 hours (Quip), Plus gets 3 hours per 24 hours (full model), Pro $100 gets 15 hours and Pro $200 is unlimited. The developer caveat: this is consumer ChatGPT Voice only at launch — OpenAI said API access “would follow” with no date or model ID. So this is a preview of the model you’ll be tuning full-duplex voice agents on, not a ship-today announcement. Voice-agent builders should also note the multilingual coverage widening significantly — it changes which languages you can prototype voice UX in first.

Quick hits and radar: what else is on the calendar

Mistral Large 4 open weights land October 27. Still on track per last week’s radar: the 1-trillion-parameter (49B active) MoE with 1M context drops under a custom Mistral licence, with the API preview live now. If you’re planning an open-weight evaluation sprint, mark the date.
Anthropic’s “cruelty clause” takes effect November 12. No new enforcement mechanism, amendment or reversal in this week’s coverage — still on track, no action needed, but know it’s coming.
Anthropic × Singapore: agentic AI evaluation MOU and a fifth APAC office. Signed at Singapore International Cyber Week on October 13 between Anthropic and Singapore’s Ministry of Digital Development and Information: cooperation on agentic AI evaluation and testing with Singapore’s AI Safety Institute, plus a new Raffles Place office. The developer angle: 1,500+ GovTech Singapore engineers already use Claude Code, and Singapore ranks 2nd of 121 countries in Claude usage per capita per the Anthropic Economic Index.
NVIDIA Bolt E4B 4GB drops October 23. The edge/conversational member of the Nemotron 3.5 open-model family — worth a look if you deploy agents to constrained hardware.
Kimi Code Desktop 1.1.0 (provisional). Moonshot’s desktop coding agent adds Linux builds (AppImage/deb/RPM) and “Remote Control” — steering the local agent from your phone via link or QR, with approvals staying on the desktop. Single source chain, so treat as provisional until it’s confirmed.

What to do this week

Three builder actions from this AI developer news cycle:

This weekend: grep your production configs, runbooks and eval harnesses for gpt-4o and gpt-4.1 model strings (including fine-tuned variants) and plan the migration to gpt-6-luna. If you’re still pinned to gpt-4o-2024-05-13 or gpt-4.1-nano, your deadline is October 23 — days, not months.
Before Monday: audit pinned models in your Copilot configs — IDE defaults, agent-mode setups and Copilot CLI scripts — for the three November 16 retirements (Claude Opus 4.8, Opus 5, GPT-5.6 Terra) and move pins to the replacements.
This week: schedule one open-weight evaluation. The shortlist writes itself: Reflection Beam (most permissive licence — Apache 2.0, strongest US open coding model) and NVIDIA Bolt 74B-A9B (the one that actually fits your 24GB card). Run the benchmark-backed model comparison framework against them before you renew your next API contract.

AI Developer News: OpenAI Retires GPT-4o, Reflection Open-Sources Beam 501B — Week Ending October 17, 2026