OpenAI's GPT-5 & Open-Weight Models, Anthropic Winning the Enterprise, Google's Genie 3, Nvidia & AMD Pay for China
OpenAI drops GPT-5 while targeting Meta with open-weight models, Anthropic's Claude Opus 4.1 is winning the enterprise, Google's world simulator impresses, and Nvidia & AMD pay 15% for China access
OpenAI's GPT-5 Launch: High Expectations, Mixed Results, and User Backlash
OpenAI finally launched GPT-5, and Sam Altman hyped it up with this post.
Expectations were high, as was the pressure after GPT 4.5 was deprecated just months after release in July.
And while GPT-5 showed steady improvement, it missed public expectations. Polymarket odds for OpenAI having the best AI model by end of 2025 crashed from 40% pre-launch to 18%, dropping them third behind xAI, while making Google the clear front-runner.
And despite CEO Sam Altman touting GPT-5’s vibe coding capabilities, Anthropic’s Claude 4 Opus is still the best, according to SWE-bench.
While Sam Altman didn’t do himself any favors with the Death Star post, GPT-5 continued the rapid progress we have seen from large language models. Super intelligence isn’t imminent, but large languages are still a big deal making rapid progress.
ChatGPT’s Backlash from Plus Users
The real story of the launch, however, quickly became OpenAI's new consumer interface, ChatGPT 5. The company introduced a redesigned model picker and a "Model Router" designed to automatically select the best model for a query's complexity. This feature was intended to make advanced reasoning more accessible, boosting usage among free users.
But the change sparked immediate backlash from Plus subscribers. Their weekly usage of advanced reasoning models dropped dramatically, from 2,900 queries to just 200. Paid users complained that the router took away the very features they were paying for. Compounding the issue, users also voiced a strong attachment to GPT-4o’s persona, with many likening it to losing a friend.
In response, OpenAI quickly brought back legacy models like GPT-4o for their paid users. Altman acknowledged the emotional connection users form with AI personas, calling it a unique challenge in product development. He also promised to restore reasoning allocations for subscribers as the company works through capacity constraints.
Ultimately, GPT-5 serves as a powerful illustration of the challenges facing AI companies. While the model made solid technical progress, it exposes OpenAI's struggles with product prioritization, managing user expectations, and the unexpected human-like bonds forming between users and AI.
With a billion users, OpenAI has to keep a lot of people happy. Upsetting just 1% of them is 10 million unhappy users, and while those unhappy users are happy about the rollbacks, how do the rest of their users feel?
OpenAI Targets Meta with First Open-Weight Models Since GPT-2
OpenAI released gpt-oss, a family of open-weight models including 120B and 20B parameter versions under a permissive Apache 2.0 license.
This marks the first time OpenAI released open-weight models since GPT-2, but shouldn’t be a big surprise. Sam Altman hinted this was coming in March 2025.
The 20B version is designed to run locally on a laptop with 16GB of VRAM, and the 120B version can run on a single H100 GPU.
There is one notable downside, which is that gpt-oss may have been trained on only synthetic data. This is understandable, as it would help avoid lawsuits that competitors like Anthropic are facing, but also means the LLM has significant knowledge gaps. Those knowledge gaps could mean gpt-oss performs better on benchmarks than on real-world tasks.
By releasing these open source models, OpenAI might be seizing Meta’s recent stumbles with their Llama models. Meta has been facing internal struggles with the Llama team, even postponing the release of Llama 4 Behemoth due to performance and development issues.
Now we’ll see how quickly Meta responds after their aggressive hiring spree. Meta has spent a fortune on GPUs and talent, how quickly can they rebound from their Llama 4 struggles?
Anthropic's Claude Opus 4.1 Takes Enterprise Crown
Anthropic announced Claude Opus 4.1 a day before OpenAI’s GPT-5 launch, and while the timing may have been a coincidence, the results were not.
The model improved agentic tasks and real-world coding by 42% over OpenAI’s prior flagship model. They are winning the enterprise thanks to their success, with 32% market share versus OpenAI's 25%.
As mentioned earlier, Claude Opus 4.1 remained the best coding tool after GPT-5’s release. GPT-5 claimed higher performance, but SemiAnalysis showed how OpenAI cherry-picked the testing by not answering 23 of the 500 SWE-bench questions.
Anthropic is making $5 billion ARR and projects $9 billion by year-end, but much of that is thanks to a few key clients, most notably the AI coding platform Cursor.
While the Cursor partnership has been highly lucrative, it’s also a major vulnerability. If Cursor switches from Claude to a competitor like Google or OpenAI, Anthropic could lose 10%-20% of their revenue overnight.
This forces Anthropic to stay in the leading edge, and could explain why they released Opus 4.1 the day before GPT-5.
Google DeepMind Launches Genie 3 World Simulator
Google DeepMind released Genie 3, an advanced simulator that generates dynamic, interactive 3D worlds from simple text prompts in real-time. This marks a major leap forward from previous models, condensing what would take years of simulation work into a matter of weeks.
The simulator allows AI agents and robots to train in physics-aware environments. This 10x improvement over prior models could revolutionize fields like robotics, gaming, and AI-for-science applications like drug discovery.
Compared to rival simulators, such as those from Meta, Genie 3 has a notable multilingual edge, broadening its potential for global access.
This release follows recent upgrades to Google’s Gemini models, solidifying our prior claims that Google is the vertically integrated player for AI.
Nvidia and AMD Agree to 15% China Revenue Share with US
Nvidia and AMD are paying 15% of China AI chip sales (H20 and MI308 models) to the US government, securing export approvals amid trade tensions. This opens up a key trade channel, but raises longer-term concerns about "pay to play" deals.
This situation echoes a prior deal where Apple pledged a $600 billion investment in the US to avoid a 25% iPhone tariff. While these deals could fund infrastructure projects, it could also create a caste of winners and losers. How would a small business, for example, make a similar pledge to avoid the tariffs they’re facing?
For businesses, this is a pragmatic short-term solution for market access, but it adds longer term uncertainty. Will this kind of deal spread to other sectors, and what happens if governments change their terms?