SI Agent News
DevelopingClaudeOpenAIGemini

LLM release cycles shorten as model capacity expands

Analysis of market leaders shows flagship model release gaps have shrunk significantly while output limits have jumped over 30x since 2023.

Risograph poster: picture for the story 'LLM release cycles shorten as model capacity expands'
AI-made risograph poster for this story. No real people or products are shown.
31xincrease in flagship output limits

The short version

Why it matters

The rapid acceleration of release cycles means developers must adapt to new model capabilities more frequently. Furthermore, the massive expansion in output limits allows for much longer, more complex generations in a single request.

Accelerating release cycles

The pace at which major AI labs release new models is intensifying. According to data analyzed by metemorris.com, the median gap between flagship model releases for OpenAI, Anthropic, and Google dropped from 119 days in 2023 to 47.5 days in 2025. This trend suggests that more resources are being poured into the development of high-capability models.

While the pace slowed slightly to a median of 57 days in 2026, the overall trend remains much faster than in previous years. OpenAI saw its median gap move from 46 to 41 days, while Anthropic's gap narrowed from 111 to 60 days. These shorter cycles mean that the state-of-the-art is shifting more rapidly than it did during the early stages of the LLM boom.

However, not all model tiers are moving at the same speed. Economy models, which are designed for low cost and fast responses, are seeing sparser releases at Anthropic and Google. This may be due to pressure from high-quality open-weight models, which provide a cheaper alternative for routine tasks and reduce the need for proprietary economy-tier offerings.

Expanding model capacities

The amount of information a model can process and generate has grown substantially since the launch of GPT-3.5. For flagship models, the median output limit—the cap on what a model can generate in one request—rose from about 4K tokens in 2023 to 128K tokens in 2026. This represents a 31x increase in the length of single-turn responses.

Input capacity, or the context window, has also seen a significant jump. The median prompt capacity for flagship models moved from 100K to 1M tokens, a 10x increase. While Google previously offered a 2M-token context window with Gemini 1.5 Pro, most recent flagship releases in 2026 have hovered around the 1M token mark, which is sufficient for many complex tasks.

These expanded limits change how developers build applications. With larger windows, agents can handle more raw material at once, though developers are also turning to techniques like compaction and agent memory. Compaction summarizes older turns to save space, while agent memory stores decisions outside the window so the model can pick up where it left off in future sessions.

The mechanics of pricing

Understanding LLM pricing requires looking at the two phases of generation: prefill and decode. Prefill processes the prompt in parallel to build a key-value (KV) cache, which is a compute-heavy task. Decode then reads and extends that cache one token at a time, a process that is often limited by memory bandwidth rather than raw compute power.

These different technical requirements explain why providers often charge different rates for input and output tokens. In the flagship tier, the median cost ratio between output and input tokens is 5x. This ratio can fluctuate based on serving efficiency and specific pricing decisions. For example, turning on 'thinking' features on Gemini 2.5 Flash Preview increased its output rate from $0.60 to $3.50 per million tokens.

Pricing trends vary significantly across different model tiers. In the economy tier, OpenAI's prices have fluctuated, while Anthropic's Haiku 5.5 saw a 10x price drop compared to its predecessor. In the balanced tier, Google's Flash prices rose from early lows, while Anthropic held its Sonnet prices steady for several generations before eventually cutting both rates.

Efficiency and research

To manage the costs and speeds of these expanding models, researchers are developing new inference methods. On the prefill side, LMCache allows inference workers to share the KV cache for repeated prompt prefixes. This allows a request to skip the prefill work entirely even if it lands on a different worker, improving speed for common prompts.

On the decode side, a method called DFlash is being used to speed up speculative decoding. This technique drafts a block of tokens in one pass for the main model to verify. The authors reported more than 6x faster generation than regular decoding in their benchmarks, helping to make long-form generation more viable for users.

These technical advancements are essential as models move toward larger output limits. As the focus shifts from simply increasing the size of the context window to getting more out of those limits, frameworks and harnesses become more important. These tools help manage the practical bottlenecks that arise when a model is tasked with processing massive amounts of information in a single session.

What to watch

Try it

Check the documentation for your preferred LLM provider to see if they use a 'shared window' or separate input/output caps for their context limits.

Sources and links

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

Listen to the AI Wire headlines

Listen to the AI Wire

Listen to our daily AI shows every morning: on your commute, in the shower, wherever you are

Listen to the daily shows

First reported by metemorris.com · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

More AI news