The AI Wire ●

xAI began rolling out Grok 4.7 to testers in mid-September 2026 after months of delays. The release continued xAI's strategy of competing on price as much as benchmarks, undercutting rivals rather than chasing outright benchmark leads.

The delays had frustrated Grok watchers, with Musk's promised timelines slipping repeatedly through 2026. The tester rollout suggested a wider release was close but not yet finalized.

Meanwhile Grok 5 — which Musk called 'AGI-class' on September 14 — remains unreleased with no date set, after missing its original Q4 2025 and Q1 2026 targets.

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.

Google released Gemini 3.7 Flash, the third Flash iteration of 2026, tuned for agentic coding chains and less verbose output. It targeted developers running multi-step agent loops where token efficiency matters.

The release continued the monthly Flash cadence with incremental but compounding gains. Google's messaging emphasized reliability over long agent runs rather than raw benchmark scores.

By August the pattern was unmistakable: Flash got the regular upgrades, Pro got the delays — and the market had largely accepted the trade.

xAI launched Grok 4.6, an agent-focused post-training upgrade on the 4.5 base with a 500,000-token context window. It tied GPT-5.6 Sol on the Artificial Analysis intelligence index while priced at just $2/$6 per million tokens — far below rivals.

The release came after SpaceX acquired xAI in February 2026, with the combined company styling itself SpaceXAI. Grok 4.6 was its first major model launch under the new structure.

Grok 4.6's pitch was value: near-frontier intelligence at budget pricing, undercutting competitors who charged multiples more for similar benchmark scores.

xAI released Grok 4.6 on August 12, 2026, a post-training upgrade on the same V9 base as Grok 4.5 pitched as its frontier model for coding, agentic tool use and knowledge work. It keeps the 500,000-token context window, accepts text and image inputs with text-only output, and offers selectable reasoning effort (low, medium, high, xhigh).

Pricing is $2 input / $6 output per million tokens for prompts under 200k tokens, doubling to $4/$12 above that line, with a faster variant at a premium. Available in the xAI API, Cursor, Grok Build, GitHub Copilot, OpenRouter, Vercel, Cloudflare and Microsoft Foundry. xAI reports gains over Grok 4.5 across its benchmark table, with independent Artificial Analysis testing placing it second only to Claude Opus 5 on agentic knowledge work at much lower token cost. (Filed September 21 as backfill.)

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.

Anthropic released Claude Opus 5, the successor flagship in the mainstream numbered line, shipping about six weeks after the premium Fable 5. Opus 5 carried forward the 4.x architecture's strengths at standard Opus pricing.

The release clarified Anthropic's two-track strategy: Fable for customers who would pay double for maximum capability, Opus 5 for everyone else. Sonnet tiers continued as the high-volume workhorse below both.

Opus 5 anchored the numbered line through the summer while Anthropic refined the Fable tier for its 5.1 update in September.

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.

Anthropic launched Claude Fable 5, a new top public tier priced at $10 per million input tokens and $50 per million output tokens — roughly double Opus pricing. Alongside it came Claude Mythos 5, the same model without the safety classifier, restricted to vetted institutions.

The launch was rocky: US export controls briefly halted Fable 5 service from June 12 to June 30, and the pricing drew complaints from developers used to cheaper flagships. Anthropic bet that top-end capability justified a luxury tier.

Fable 5 split Anthropic's lineup in two — a premium Fable line and the mainstream numbered Opus line — a structure the company kept through the 5.1 refresh in September.

Anthropic released Claude Opus 4.8, its mid-2026 flagship for reasoning, coding, and autonomous agents. The release continued the rapid 4.x cadence, landing three months after Opus 4.6.

Opus 4.8 refined long-horizon agentic reliability and reasoning depth rather than introducing a new architecture. Anthropic was by then shipping meaningful upgrades roughly quarterly.

It was the last 4.x flagship before Anthropic's lineup split into the premium Fable tier and the numbered Opus 5 line in the summer of 2026.

Google released Gemini 3.5 Flash, starting an accelerating cadence of roughly one Flash release a month through 2026. The company was clearly prioritizing its cheap, fast tier while the next Pro flagship kept slipping.

Each Flash iteration brought small but real gains in agentic coding and instruction following, keeping Google competitive on price-performance even without a new Pro. Developers got a steady drip of upgrades.

The Flash treadmill became the defining pattern of Google's 2026: four Flash releases shipped while the long-awaited Gemini 3.5 Pro never materialized.

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.

After CEO Dario Amodei refused the Department of War's demand to drop safeguards barring Claude's use for mass surveillance of Americans or fully autonomous weapons — a $200M contract clause the Pentagon wanted rewritten to permit "any lawful use" — President Trump posted on Truth Social ordering every federal agency to stop using Anthropic technology, with a six-month phaseout, and warned he would use "the Full Power of the Presidency to make them comply, with major civil and criminal consequences to follow." Defense Secretary Hegseth separately designated Anthropic a "supply chain risk," a label never before applied to an American company.

Amodei called the threats retaliatory and said Anthropic would challenge the designation in court; a judge later blocked the ban as illegal First Amendment retaliation.

Google previewed Gemini 3.1 Pro, reporting 77.1% on ARC-AGI-2 — roughly double Gemini 3 Pro's score — and 80.6% on SWE-bench Verified. The preview suggested Google's reasoning gains were still accelerating.

ARC-AGI-2, designed to resist benchmark saturation, had humbled every previous flagship; 3.1 Pro's jump was one of the largest single-generation gains on the test. Google framed it as evidence its Deep Think approach had headroom.

The preview kept Google in the 2026 frontier conversation even as the company prioritized shipping Flash iterations over finalizing the next full Pro launch.