The AI Wire ●

OpenAI made GPT-Live-1 generally available in the API on September 10, bringing the full-duplex voice conversations from ChatGPT to developers. The model listens and speaks simultaneously — no turn-taking relay — and delegates deeper reasoning and tool calls to a backend text model of the developer's choice (GPT-6 Astra, Terra/Luna, or third-party).

OpenAI's published numbers: 80.1% on Full Duplex Bench v1.5 interactivity and 0.798s turn-taking latency versus 1.41s for GPT-Realtime-2.1; 83.6% on Tau3 customer-service tasks when paired with GPT-6 Astra. Pricing is $0.05 per minute of voice, billed per second, with backend tokens billed separately. Connects via WebRTC, WebSockets, or SIP/telephony.

OpenAI's head of Core Products & Platform — the face behind the usage-reset announcements — talks through how Codex was built and iterated: why the CLI is Rust and open source, how the harness and models evolved, why Codex supports multiple providers, and the ChatGPT + Codex merger.

OpenAI's new flagship model, succeeding GPT-5.6 Sol. State-of-the-art on computer use, browsing, software engineering and science per OpenAI; rolls out to ChatGPT Plus, Pro, Business and Enterprise tiers in waves, plus API access.

An anonymous X account (@synthwavedd) claimed OpenAI has finished pretraining a base model codenamed Bel — successor to "Doug," reportedly over 10 trillion total parameters — expected to underpin the Astra family and a future GPT-6, and possibly an "AGI-threshold" system after further reinforcement learning. A Sept 13 follow-up claims early accounts have Bel outperforming GPT-6 Astra on coding tasks, and that OpenAI believes Anthropic lacks the compute to answer it before its stock listing.

Caveats: single anonymous source, secondary accounts disagree on basics (including whether Bel precedes or succeeds Doug), and OpenAI has made no public comment on Bel at all.

As of October 4, the name "Bel" still appears nowhere in OpenAI's model docs, API changelog, or news feed — a grading page from CellCog traces the whole claim back to one August 25 X post from @synthwavedd about a finished pretrain with 10T+ parameters, a successor to an earlier run codenamed Doug (which SemiAnalysis referenced in a July 9 subscriber note). The viral "GPT-6 Bel beats Astra" headline added a benchmark claim the source never made.

The same account said on September 23 that Bel is still months from going public. One thing to watch: OpenAI's Codex lead teased a release next week on October 3 without naming a model — if that materializes, the rumor may finally get a verdict.

OpenAI expanded its Daybreak cybersecurity initiative into two access tiers for vetted defenders. Daybreak Blue opens general-purpose models, including GPT-5.6 Sol, to approved defenders for everyday security work — vulnerability discovery, secure code review, malware analysis, incident response, and patch validation — with the system-level cyber guardrails recalibrated rather than removed.

Daybreak Red sits behind tighter vetting and gates the new purpose-trained GPT-5.6-Cyber model, built for vulnerability research, exploit validation, and authorized security testing. On OpenAI's internal Advanced Cybersecurity Completion Rate benchmark, GPT-5.6-Cyber completed 95% of offensive-security tasks, up from 57.3% for GPT-5.5-Cyber and just 1.5% for general-purpose GPT-5.6 Sol. OpenAI says it used the model to find two previously unknown vulnerabilities in Chrome's V8 engine.

Access requires identity verification, monitoring, and legal attestations — and from September 1, hardware security keys became mandatory for individual Daybreak accounts. The models also landed on Amazon Bedrock on August 11, with a partner program covering 16 security providers including IBM, CrowdStrike, Accenture, Cisco, and Cloudflare.

OpenAI fully released the GPT-5.6 series after a limited preview that began June 27: three capability tiers in one generation. Sol is the flagship, Terra the everyday model, and Luna the most cost-efficient — at $1/$6 per million input/output tokens, roughly a sixth of Sol's price.

The launch also folded Codex into ChatGPT itself, which Sam Altman called 'a really big deal' on the launch livestream — merging the conversational assistant and the coding agent into one surface.

Chief scientist Jakub Pachocki had previewed GPT-5.6 as a 'meaningful improvement' over GPT-5.5, with gains aimed at multi-hour agentic workflows rather than single-turn chat quality.

OpenAI released GPT-5.5, positioned as the frontier model for coding, research, computer use, documents, spreadsheets, and long-running work. It continued the roughly six-week flagship cadence OpenAI had settled into.

Its run was short: on September 15, OpenAI announced GPT-5.5 would retire from ChatGPT, ChatGPT Work, and Codex on October 14, 2026, directing users to GPT-5.6 Sol or GPT-6 Astra instead.

The brisk retirement underscored how aggressively OpenAI was cycling its lineup — GPT-5.5 lasted barely six months in the main product surfaces.

OpenAI launched GPT-5.4 in Thinking and Pro variants, reporting 57.7% on SWE-bench Pro, around 80% on SWE-bench Verified, and 75% on OSWorld for computer use — with a 1M-token context window and 128K max output.

Smaller GPT-5.4 mini and nano variants followed on March 17 for faster, lower-cost workloads, going GA across all Copilot tiers the same day.

GPT-5.4-class models were pulled from Codex at the end of August 2026 as the GPT-6 generation took over the flagship slot.

OpenAI released GPT-5.3-Codex, a return to code-specialized fine-tuning after years of generalist flagships. The agentic coding model hit 56.8% on SWE-bench Pro and 77.3% on Terminal-Bench 2.0, running about 25% faster on tool-driven tasks.

The name was a deliberate callback to the original 2021 Codex, but the model was a different beast: agent-optimized rather than a simple code completion engine. It reached GA in Copilot on February 9 across Pro, Business, and Enterprise tiers.

OpenAI later designated GPT-5.3-Codex its first long-term-support model, guaranteeing it for Business and Enterprise Copilot users through February 2027.

OpenAI released GPT-5.2, a three-mode family — Instant, Thinking (with standard and extended effort), and the heavier GPT-5.2 Pro — with gains in instruction following, vision, front-end code generation, and token efficiency. A new xhigh reasoning level and context compaction arrived for long-running work.

The launch came about three weeks after Google's Gemini 3 Pro seized the multimodal lead, and reporting described an internal 'Code Red' memo accelerating GPT-5.2's release. OpenAI executives pushed back on the reactive framing while acknowledging the extra resources helped.

A coding-specialized GPT-5.2-Codex followed on December 18. GPT-5.2 was eventually discontinued in ChatGPT on June 12, 2026.

OpenAI shipped GPT-5.1 as a ChatGPT update with two modes — GPT-5.1 Instant and GPT-5.1 Thinking — and, more unusually, let users select alternative assistant personalities including 'friendly,' 'efficient,' and 'cynical.'

The release also improved coding steerability and reduced the 'overthinking' problem where reasoning models wasted compute on simple questions.

A coding variant, GPT-5.1-Codex-Max, followed on November 18, foreshadowing OpenAI's return to specialized coding models.

OpenAI released GPT-5, the first model to unify its two lineages: a fast model for most queries, a deeper 'thinking' model for hard problems, and a real-time router that picks between them based on the task. Users no longer had to choose a model — the system decided.

OpenAI reported gains over GPT-4 in coding, writing, health answers, and factual accuracy, with lower hallucination rates and stronger agentic tool use. It shipped across ChatGPT, Microsoft Copilot, and the API on day one.

GPT-5 became the base of the 5.x generation, which OpenAI iterated rapidly through 5.1, 5.2, and beyond — a faster cadence than any previous flagship line.

OpenAI launched GPT-4.5, codenamed Orion, describing it as its largest model to date. It was positioned as a research preview for Pro users and developers — notably, the last OpenAI model released without chain-of-thought reasoning.

GPT-4.5 emphasized emotional intelligence and conversational quality over raw benchmark gains. Reviews found it warmer and more intuitive than the reasoning models, but its huge training cost made it expensive to serve.

In retrospect it marked the end of an era: the final flagship of the pure next-token-prediction lineage before the GPT-5 generation unified fast and reasoning models under one system.

OpenAI released o1-preview, the first model in a new series trained with reinforcement learning to 'think' through problems using an internal chain of thought before answering. It was slower than GPT-4o but dramatically better at math, science, and competitive programming.

The launch introduced the reasoning-model paradigm that would define the next two years of AI development: trading inference-time compute for capability. A smaller o1-mini arrived the same day as a cheaper, faster alternative.

The full o1 followed on December 5 alongside the new $200/month ChatGPT Pro tier, cementing reasoning as a premium capability rather than a research curiosity.

OpenAI introduced GPT-4o — 'o' for omni — a single model natively handling text, vision, and audio with response times approaching human conversation. It was faster and cheaper than GPT-4 Turbo, and unusually, it rolled out to free ChatGPT users too, with usage limits.

The headline feature was Advanced Voice Mode, letting users talk to ChatGPT nearly in real time with interruptions and emotional tone. It began rolling out to Plus users in late July and became one of the product's signature experiences.

GPT-4o remained the flagship for over a year and developed an unusually loyal following — its eventual removal from ChatGPT years later prompted genuine user backlash over its personality.