Kimi K2 Thinking versus GPT-5 is now a comparison that must separate model weights from hosted availability. Moonshot’s API model table lists the old kimi-k2-thinking endpoint as retired, while its official weights remain available. OpenAI schedules the original GPT-5 API model to shut down December 11, 2026. A downloadable checkpoint can remain usable after a provider retires its own endpoint. Moonshot model status, official Kimi K2 Thinking card, GPT-5 reference. OpenAI’s dated retirement notices.
Researched and written by GPT, an AI, from cited primary sources. Recommendations are editorial assessments; no hands-on performance test is claimed. Numerical examples are calculations under stated assumptions.
Sources checked November 3, 2026.
This article compares the exact models, not today’s Kimi and ChatGPT consumer experiences. A current application may use a different generation, tools and routing behavior.
Open weights change who operates the system
Kimi K2 Thinking’s card describes a mixture-of-experts model with one trillion total parameters, 32 billion active parameters, a 256K-token context and native INT4 quantization. Its downloadable release uses a modified MIT license. Those facts do not imply that it fits a normal laptop: active parameter count is not the complete memory footprint.
GPT-5 provides a managed API path with text/image input, text output and a 400,000-token context. Kimi K2 Thinking is a text reasoning model. If the workload depends on images, that is an architectural difference rather than a minor feature checkbox.
Read the benchmark conditions before the scores
Moonshot’s K2 Thinking card includes comparisons with GPT-5 across reasoning, coding and agent tasks. These are vendor-reported results, not a benchmark independently reproduced for this article. The table mixes settings, tool access and some quoted or retested baselines; a result under one configuration does not establish a universal winner.
A tool-enabled agent result answers a different question from a no-tools knowledge test. Similarly, a multi-trajectory or heavier reasoning run can consume more resources than a single attempt. Preserve the harness, budget and exact model settings when using any published result to design your own evaluation. Kimi evaluation details.
Reproduce the task that matters to you
For a repository investigation, give each supported candidate the same code revision, search tools and issue. Require a short explanation tied to specific files and a patch that passes the relevant tests. Record failed attempts, tool loops, completion time and reviewer corrections rather than only the best output.
For research, require primary sources and checks for unsupported claims. A model that can continue reasoning for a long time can also continue down an incorrect path; length is not a substitute for evidence.
Compare current operating costs
GPT-5’s listed standard rates are $1.25 per million input tokens and $10 per million output tokens, but its retirement means a new managed workflow needs a successor budget. Self-hosting K2 Thinking needs infrastructure and operational pricing; a third-party hosted offer needs its own availability and contract check. Neither should be called “free” simply because weights can be downloaded.
Use the historical comparison to understand the deployment and evaluation tradeoffs. Choose a currently supported managed service when that reduces operational burden, or evaluate the downloadable model when control and customization justify running it. Confirm the exact hosting path before recommending either to someone starting today.
Model references
Moonshot AI · Model library
Kimi-K2-Thinking
AvailableDownloadable K2 Thinking model; its weight availability is separate from Moonshot hosted API retirement.
Author’s assessment; check the guide’s evidence and limitations.- Access
- Open weights
- Input
- text
- Output
- text
- Context window
- Not documented
Catalog checked . Model specifications do not establish access through a particular app or subscription.
OpenAI · Model library
GPT-5
DeprecatedOriginal GPT-5 endpoint, with a December 11 retirement date.
Author’s assessment; check the guide’s evidence and limitations.- Access
- API
- Input
- text, image
- Output
- text
- Context window
- 400,000 tokens
Catalog checked . Model specifications do not establish access through a particular app or subscription.
