aillmsinferenceresearch

Tokenomiks I: LLM Releases, Limits, and Pricing

November 14, 2026 • 8 min read

Since ChatGPT launched in November 2022, powered by GPT-3.5, the software industry has rapidly moved LLMs to center stage. For many software engineers, the job now involves LLMs in some way, whether that’s writing code with them, developing the models themselves, building the infrastructure they run on, or integrating them into products.

But since LLMs became mainstream, how have the models themselves changed? It is easy to get caught up in each new release without realizing how far we have come.

As in site reliability engineering, I wanted to start by looking at the metrics. As the first part of this series, this post is focused on model release frequency, token limits, and token pricing.

I wanted this initial analysis to focus on market leaders, so I looked at major proprietary model releases from five US labs through November 11, 2026. Most charts start with OpenAI, Anthropic, and Google. You can also select Meta and SpaceX/xAI.

Grouping Models

In order to avoid comparing apples to oranges, I grouped models by how the labs positioned them.

  • Economy: Models built for low cost and fast responses, especially when an application needs to make a lot of calls.

    Example: Claude Haiku 5.5

    View all 40 models

    OpenAI

    • codex-mini-latest (initial model)
    • gpt-3.5-turbo-0125
    • gpt-3.5-turbo-0301
    • gpt-3.5-turbo-0613
    • gpt-3.5-turbo-1106
    • gpt-3.5-turbo-16k-0613
    • gpt-4.1-mini-2025-04-14
    • gpt-4.1-nano-2025-04-14
    • gpt-4o-mini-2024-07-18
    • GPT-5 Instant Mini
    • gpt-5-mini-2025-08-07
    • gpt-5-nano-2025-08-07
    • gpt-5.1-codex-mini
    • GPT-5.3 Instant Mini
    • gpt-5.4-mini-2026-03-17
    • gpt-5.4-nano-2026-03-17
    • GPT-5.5 Instant Mini
    • gpt-5.6-luna
    • gpt-6-luna
    • o1-mini-2024-09-12
    • o3-mini-2025-01-31
    • o4-mini-2025-04-16

    Anthropic

    • Claude Instant (March 2023 launch)
    • claude-3-5-haiku-20241022
    • claude-3-haiku-20240307
    • claude-haiku-4-5-20251001
    • claude-haiku-5-5
    • claude-instant-1.2

    Google

    • gemini-1.5-flash-8b
    • gemini-1.5-flash-8b-exp-0827
    • gemini-1.5-flash-8b-exp-0924
    • gemini-2.0-flash-lite
    • gemini-2.0-flash-lite-preview-02-05
    • gemini-2.5-flash-lite
    • gemini-2.5-flash-lite-preview-06-17
    • gemini-2.5-flash-lite-preview-09-2025
    • gemini-3.1-flash-lite
    • gemini-3.5-flash-lite

    Meta

    • muse-ember-1.0

    SpaceX / xAI

    • grok-3-mini

  • Balanced: Everyday models that aim for a practical mix of capability, speed, and cost.

    Example: Claude Sonnet 5.5

    View all 41 models

    OpenAI

    • GPT-5 Instant
    • GPT-5.1 Instant
    • GPT-5.2 Instant
    • GPT-5.3 Instant (initial gpt-5.3-chat-latest)
    • GPT-5.5 Instant
    • gpt-5.6-terra
    • gpt-6.1-terra

    Anthropic

    • claude-3-5-sonnet-20240620
    • claude-3-5-sonnet-20241022
    • claude-3-7-sonnet-20250219
    • claude-3-sonnet-20240229
    • claude-sonnet-4-20250514
    • claude-sonnet-4-5-20250929
    • claude-sonnet-4-6
    • claude-sonnet-5
    • claude-sonnet-5-5

    Google

    • chat-bison@001
    • chat-bison@002
    • code-bison@002
    • codechat-bison@002
    • gemini-1.5-flash
    • gemini-1.5-flash-001
    • gemini-1.5-flash-002
    • gemini-1.5-flash-exp-0827
    • gemini-2.0-flash
    • gemini-2.0-flash-exp
    • gemini-2.0-flash-thinking-exp-01-21
    • gemini-2.0-flash-thinking-exp-1219
    • gemini-2.5-flash
    • gemini-2.5-flash-preview-04-17
    • gemini-2.5-flash-preview-05-20
    • gemini-2.5-flash-preview-09-2025
    • gemini-3-flash-preview
    • gemini-3.5-flash
    • gemini-3.6-flash
    • gemini-3.7-flash
    • gemini-3.8-flash
    • text-bison@001
    • text-bison@002

    SpaceX / xAI

    • grok-4-1-fast-non-reasoning
    • grok-4-1-fast-reasoning

  • Flagship: A company’s main high-capability models, intended for more demanding workloads.

    Example: Claude Opus 5.5

    View all 69 models

    OpenAI

    • codex-1
    • gpt-4-0125-preview
    • gpt-4-0314
    • gpt-4-0613
    • gpt-4-1106-preview
    • gpt-4-32k-0314
    • gpt-4-32k-0613
    • gpt-4-turbo-2024-04-09
    • gpt-4.1-2025-04-14
    • gpt-4.5-preview-2025-02-27
    • gpt-4o-2024-05-13
    • gpt-4o-2024-08-06
    • gpt-4o-2024-11-20
    • gpt-5-2025-08-07
    • gpt-5-codex
    • gpt-5.1-2025-11-13
    • gpt-5.1-codex
    • gpt-5.1-codex-max
    • gpt-5.2-2025-12-11
    • gpt-5.2-codex
    • gpt-5.3-codex
    • gpt-5.4-2026-03-05
    • gpt-5.5-2026-04-23
    • gpt-5.6-sol
    • gpt-6-sol
    • gpt-6.1-sol
    • o1-2024-12-17
    • o1-preview-2024-09-12
    • o3-2025-04-16

    Anthropic

    • Claude (March 2023 launch)
    • claude-2.0
    • claude-2.1
    • claude-3-opus-20240229
    • claude-opus-4-1-20250805
    • claude-opus-4-20250514
    • claude-opus-4-5-20251101
    • claude-opus-4-6
    • claude-opus-4-7
    • claude-opus-4-8
    • claude-opus-5
    • claude-opus-5-5

    Google

    • gemini-1.0-pro
    • gemini-1.5-pro
    • gemini-1.5-pro-001
    • gemini-1.5-pro-002
    • gemini-1.5-pro-exp-0801
    • gemini-1.5-pro-exp-0827
    • gemini-2.0-pro-exp-02-05
    • gemini-2.5-pro
    • gemini-2.5-pro-exp-03-25
    • gemini-2.5-pro-preview-03-25
    • gemini-2.5-pro-preview-05-06
    • gemini-2.5-pro-preview-06-05
    • gemini-3-pro-preview
    • gemini-3.1-pro-preview
    • text-unicorn (initial GA)

    Meta

    • Muse Spark
    • Muse Spark 1.1
    • Muse Spark 1.2
    • Muse Spark 1.3
    • Muse Wonder

    SpaceX / xAI

    • grok-1-product
    • grok-2-beta
    • grok-3
    • grok-4-0709
    • grok-4.20
    • grok-4.5
    • grok-4.6
    • grok-4.7

  • Premium: Models positioned above the regular flagship, with an emphasis on pushing capability further.

    Example: Claude Fable 5.2

    View all 13 models

    OpenAI

    • gpt-5-pro
    • gpt-5.2-pro
    • gpt-5.4-pro
    • gpt-5.5-pro
    • gpt-6-astra
    • o1-pro
    • o3-pro-2025-06-10

    Anthropic

    • claude-fable-5
    • claude-fable-5-1
    • claude-fable-5-2

    Google

    • gemini-4-argon
    • gemini-4-carbon

    Meta

    • Muse Wonder Full

How Fast Are the Models Being Released?

Loading sourced observations…

  • Flagship gaps are getting shorter. For OpenAI, Anthropic, and Google, the median fell from 119 days in 2023 to 47.5 days in 2025. It rose to 57 days in 2026. OpenAI’s median went from 46 to 41 days and Anthropic’s from 111 to 60, while Google had one flagship release this year, with Gemini 4 Argon/Carbon in Premium.
  • Economy releases are getting sparser at Anthropic and Google. Anthropic’s last two gaps were 345 and 357 days, up from 217 and 236 days in 2024. Google went from five releases in 2025 to two so far in 2026. OpenAI went from six to five.
  • Google had the shortest gaps in Balanced. In 2026, its median gap was 43 days, compared with about 74 at OpenAI and 133 at Anthropic.

My guess is that better open-weight models are putting pressure on the economy tier. If one can run routine tasks on a cheap open-weight model, people have less reason to use the potentially more expensive proprietary one.

How Much Can Models Handle?

The capacity charts show each model’s documented standard API limits, plotted by model release date. The dashed line is the highest limit released so far. Filled points were documented within 30 days of release. Hollow points come from later documentation, mostly the current docs, so the launch-day limit may have been different.

Input token limits cap what you send, including instructions, chat history, and documents. Some Gemini models publish separate input and output caps so you can check each budget up front. Claude also documents a shared window. In a 100K shared window, an 80K prompt leaves at most 20K for generation. The chart shows standalone input caps as circles and shared windows as diamonds.

Loading sourced observations…

The context window is how much information the API lets the model work with at once. In a shared window, that includes the prompt, its history, and the tokens generated on that turn.

Google offered a 2M-token context window with Gemini 1.5 Pro but no release has approached that since 2025. For flagship models, 10 of the 12 context windows for 2026 releases were about 1M tokens, up from 4 of 16 in 2025.

It seems like 1M is enough for most tasks today if the agent runs in a good harness. Compaction summarizes older turns to make room for the session to continue. Agent memory saves things like decisions and progress outside the window so the next session can pick up where the last one left off. If the harness keeps the original transcript, it can pull details back in as needed. A bigger window still helps when a task needs a lot of raw material in view at once.

Output token limits cap what the model generates in one request. On some APIs, internal reasoning and the visible answer draw from that same budget. Providers have to leave room for long answers and reasoning while keeping response times, serving capacity, and per-request costs manageable.

Loading sourced observations…

In the 2026 data, the documented standard output limits for the three labs fall into two bands. Anthropic and OpenAI are at 128k tokens, and Google is at 65k. Gemini 4 can generate beyond that default using long decode.

How Is Token Pricing Changing?

Prefill processes the prompt in parallel and builds the key-value (KV) cache. Decode reads and extends that cache as it generates one token at a time. In aggregated serving, the two phases share GPUs. In disaggregated serving, separate GPUs handle each phase and pass the cache from prefill to decode.

Prefill is typically compute-heavy, while decode is often limited by memory bandwidth. That difference helps explain why providers charge different rates for input and output.

The prices below are list prices at API launch. They don’t include later price cuts, caching or batch discounts, or how many tokens a task actually uses.

Loading sourced observations…

Input and output prices have moved differently across the four tiers. In Economy, OpenAI’s prices are lower overall but moved around a lot between releases. Anthropic’s rose from Claude 3 Haiku to Haiku 4.5, then dropped 10× with Haiku 5.5. GPT-6 Luna and Haiku 5.5 both launched at $0.10 input and $0.50 output per million tokens, which fits the open-weight pressure I mentioned above.

In Balanced, Google’s Flash prices rose from their early lows, while Anthropic held Sonnet prices steady for several generations before cutting both rates.

Flagship input and output prices are well below their early levels at OpenAI and Anthropic. Google’s started much lower and rose.

How much more does a generated token cost? The median flagship output/input cost ratio in the data is 5×. The ratio can rise even as both prices fall. GPT-4 launched at $30 input and $60 output per million tokens (2×). GPT-5 launched at $1.25 and $10 (8×). GPT-6.1 Sol was back at 5×.

Both serving efficiency and pricing decisions can move the ratio. At launch, Gemini 2.5 Flash Preview cost $0.15 per million input tokens, but turning thinking on took the output rate from $0.60 to $3.50. The same model went from 4× to about 23× depending on the thinking level.

There is a lot of ongoing research around making inference cheaper, both in prefill and decode phases. On the prefill side, LMCache lets inference workers share the KV cache for repeated prompt prefixes, so a request can skip prefill work even when it lands on a different worker. On the decode side, DFlash speeds up speculative decoding by drafting a block of tokens in one pass for the main model to verify. The authors reported more than 6× faster generation than regular decoding in their benchmarks.

Conclusions

Loading sourced observations…

Release gaps are generally getting much shorter. That is expected given more resources go into LLMs, though the pace varies across labs and model groups. Anthropic picked up the pace on both flagship and premium releases, and OpenAI’s flagship releases got faster in 2026. Google had the fastest release cadence for Balanced models in 2026, where Flash models sit.

The biggest change is how much room models have to work. For flagship models, the median output limit went from about 4K tokens in 2023 to 128K in 2026, about 31×, and the median prompt capacity went from 100K to 1M, 10×. I fondly remember the release of Gemini 2.5 Pro, when the output limit jumped from 8K to about 65K and made a bunch of my early Detectr hacks obsolete overnight.

While these limits rose rapidly during 2024 and 2025, in 2026 the focus shifted toward getting more out of those limits. Frameworks, harnesses, and memory ease some of the practical bottlenecks those limits created.

It is harder to identify a uniform trend for pricing, as it varies by model tier. In Flagship, the median launch price fell sharply after 2023, then rose again in 2026. In Economy, input launch prices fell much faster than output prices. In Balanced, both medians ended about a third below their 2024 level.

I might go deeper into some of these areas in future posts. I made all of these graphs interactive so you can group, filter, and draw your own conclusions. I’ve shared the data and methodology so you can do your own analysis.