AI Roundup November 2023: OpenAI's GPT-4 Turbo, Altman Fired Then Rehired, and Claude 2.1's 200K Window

TL;DR

November 2023 was the most chaotic month in AI's short modern history. OpenAI opened with a triumphant DevDay (GPT-4 Turbo, a 128K context window, the Assistants API, and a GPT Store), then detonated when its nonprofit board fired Sam Altman on a Friday and reinstated him five days later after almost the entire staff threatened to walk. Anthropic used the turmoil to ship Claude 2.1 with a 200K context window, and 28 governments plus the EU signed the Bletchley Declaration at the first global AI Safety Summit. Oh, and Elon Musk's xAI shipped Grok.

OpenAI DevDay: GPT-4 Turbo, 128K Context, and the GPT Store

On November 6, OpenAI held its first developer conference and dropped the kind of update that resets everyone's roadmap. GPT-4 Turbo arrived with a 128K context window (roughly 300 pages in a single prompt), a knowledge cutoff bumped to April 2023, and pricing that undercut the old GPT-4 by 3x on input tokens and 2x on output. Vision, DALL-E 3, and text-to-speech all landed in the API the same day.

The bigger story for builders was the Assistants API and GPTs. The Assistants API gave you persistent threads, built-in Code Interpreter, retrieval, and function calling, which is OpenAI quietly absorbing a chunk of what the LangChain-and-vector-DB crowd had been gluing together by hand. GPTs let anyone build a custom ChatGPT with no code, with a GPT Store promised before year end. If you ran a thin wrapper startup, this was the day to check whether your moat survived.

Why it matters

Cheaper tokens plus a 128K window plus native tool-calling means a lot of RAG plumbing got commoditized overnight. Great if you build products, uncomfortable if your product was the plumbing.

The OpenAI Board Fired Sam Altman, Then Rehired Him Five Days Later

Eleven days after DevDay, on November 17, OpenAI's board abruptly removed Altman as CEO, saying it "no longer has confidence in his ability to continue leading OpenAI." Greg Brockman quit in solidarity hours later. Microsoft, OpenAI's largest backer, reportedly got about a minute of warning.

What followed was the fastest corporate-governance meltdown in tech memory. Satya Nadella announced Microsoft would hire Altman and Brockman to run a new in-house AI lab. Then roughly 700 of OpenAI's ~770 employees signed a letter threatening to follow them unless the board resigned and Altman returned. By November 22, Altman was reinstated with a reconstituted board chaired by former Salesforce co-CEO Bret Taylor. The board never produced a concrete explanation.

Why it matters

The whole episode was a live stress test of the "capped-profit controlled by a safety nonprofit" structure, and the structure lost. If you build on a single vendor, this was a vivid reminder that your dependency chain includes their boardroom. Plenty of teams quietly added a second model provider that week.

Anthropic Ships Claude 2.1 With a 200K Context Window

On November 21, in the middle of the OpenAI circus, Anthropic shipped Claude 2.1. The headline was a 200K token context window (about 150,000 words or 500-plus pages), nearly double GPT-4 Turbo's fresh 128K. Anthropic also claimed a 2x reduction in hallucinated false statements versus Claude 2.0, added system prompts, and introduced a beta tool-use feature.

The timing was not subtle. While OpenAI's leadership was a coin flip, Anthropic put out a serious release and reminded everyone that the frontier had more than one lab on it. The 200K window was genuinely useful for long-document work, even if independent testers quickly noted that recall in the deep middle of that window was uneven.

Why it matters

For anyone doing long-context work (codebases, contracts, research dumps), having a credible second vendor with a bigger window was exactly the diversification the Altman week argued for.

Bletchley Declaration: 28 Countries Sign the First Global AI Safety Pact

On November 1 and 2, the UK hosted the first global AI Safety Summit at Bletchley Park, the old Enigma codebreaking site. The result was the Bletchley Declaration, a non-binding agreement signed by 28 countries plus the EU. Notably, both the US and China signed the same document, alongside the UK, France, Germany, Italy, and others.

The declaration commits signatories to cooperate on identifying frontier AI safety risks and building risk-based policies. It is aspirational rather than enforceable, but getting Washington and Beijing to sign the same page on AI risk was the actual achievement.

Why it matters

Regulation is coming, and this was the diplomatic on-ramp. For builders, the signal is that frontier-model providers will face safety-testing and transparency expectations, which trickles down into the terms and capabilities you get to ship on.

Elon Musk's xAI Launches Grok

Not to be left out of the busiest month of the year, Elon Musk's xAI unveiled Grok, a chatbot with a deliberately snarky personality and live access to posts on X. It rolled out first to X Premium+ subscribers. The model itself, Grok-1, was not at the GPT-4 frontier, but the pitch was real-time data and fewer guardrails.

Why it matters

Grok mattered less as a model and more as a statement: another well-funded lab with its own compute, its own data firehose, and a distribution channel baked into a social network. The frontier club kept getting more crowded.

Key Takeaways

• Token economics shifted under everyone's feet. GPT-4 Turbo made long-context, multimodal calls dramatically cheaper, so re-cost your pipelines before you optimize them.
• Vendor concentration is a real risk. The Altman firing showed your favorite API can wobble for reasons that have nothing to do with the tech. Wire in a fallback provider.
• Long context went mainstream. With 128K from OpenAI and 200K from Anthropic in the same month, whole-document workflows stopped being exotic, but test recall before you trust it.
• The platform layer is eating the wrapper layer. Assistants API and GPTs absorbed a lot of glue-code value, so build on top of the new primitives instead of reselling them.
• Policy is no longer hypothetical. Bletchley got the US and China to sign the same safety document, which means compliance expectations will eventually reach the models you depend on.