In 30 seconds
Matthew Berman says the speed of AI models matters alongside output quality, especially when agents work without a person reading each token. He cites Claude Opus 4.6 Fast Mode and GPT-5.3-Codex-Spark as examples.
- 1A case for speedBerman has long valued model speed and argues that it matters alongside quality.
- 2Agents and reading speedAn agent can act on model output without waiting for a person to read it, making speed important for its workflow.
- 3Two fast optionsBerman points to Opus 4.6 Fast Mode and GPT-5.3-Codex-Spark as examples of the trend.
Takeaways
- Berman argues that model speed becomes especially important when agents perform work without waiting for a human to read their output.
- Anthropic says Opus 4.6 Fast Mode generates output 2.5 times faster than its regular mode.
- OpenAI reported more than 1,000 tokens per second for GPT-5.3-Codex-Spark on Cerebras hardware.
- Token generation speed is only one factor in total task time and does not by itself demonstrate quality or reliability.
Read the edited text by chapter
Why speed matters
Berman argues that fast output is especially useful for agents doing real-world work. He contrasts this with a focus on model quality alone.
I feel vindicated: I have been a “speed maxi” for AI models for a long time. People have often argued that quality matters most, but I think speed is just as important. When agents can do real-world work without waiting for a human to read every output, faster generation can make a meaningful difference.
Faster model releases
He cites Anthropic’s Fast Mode and OpenAI’s Codex-Spark. The published speed figures describe output generation, not necessarily total task duration.
I have been using Claude Opus 4.6 Fast Mode and find it extremely fast. [Anthropic says](https://www.anthropic.com/research/measuring-agent-autonomy) that this mode generates output 2.5 times faster than regular Opus 4.6. This is a comparison of output generation speed, not necessarily a 2.5-fold improvement in every complete task.
OpenAI has also released GPT-5.3-Codex-Spark, which it says can generate more than 1,000 tokens per second on Cerebras hardware. Its [February 2026 announcement](https://openai.com/index/introducing-gpt-5-3-codex-spark/) described a smaller coding model offered as a research preview, initially to ChatGPT Pro users. Actual task time also depends on input processing, tool execution and network overhead; speed alone does not establish output quality. Speed maxis unite.
More from this channel
34:21EN
Matthew Berman· Oct 27, 2026Matthew Berman Examines Anthropic’s IPO Filing, AI Growth and GovernanceMatthew Berman, Sam Altman · Forward FutureThe presenter reviews the S-1’s financial metrics, compute strategy and voting structure while questioning the tension between rapid AI development and safety.
0:26ENMatthew Berman· Oct 25, 2026Sam Altman on Human Values and Problem-Solving in an AI WorldSam Altman · OpenAIIn a short interview excerpt, Altman considers the enjoyment of solving hard problems and the human role in interpreting results.
0:36ENMatthew Berman· Oct 25, 2026Kimi K3.1 and a Rubik’s Cube Solver DemonstrationMatthew Berman · Forward FutureMatthew Berman shows a scrambled virtual cube being solved and gives the Kimi K3.1 test a favorable assessment.
0:21ENMatthew Berman· Oct 24, 2026An AI-Generated Kimi Benchmark Chart Used Invented ScoresMatthew Berman · Forward FutureMatthew Berman asks for a graph of Kimi K3.1 Pro benchmarks, then discovers that the displayed figures are illustrative placeholders.
25:50ENMatthew Berman· Oct 24, 2026Kimi K3.1 Pro Tested: Strong 3D Demos, Uneven Results and Competitive PricingMatthew Berman · Forward FutureMatthew Berman tests the open-weight model on coding, 3D rendering and visual explanations, then compares benchmark results and prices.
32:20ENMatthew Berman· Oct 23, 2026Who Is Local AI For? The Costs, Trade-offs and a Hybrid FutureMatthew Berman · Forward FutureMatthew Berman weighs data control and independence against hardware cost, setup complexity and inference speed.