Both retire on April 30, 2027. Mistral names Large 4 as the replacement.
Mistral marked Mistral Medium 3.5 and Mistral Large 3 as deprecated on October 31, 2026, according to the deprecated models table on its models overview page. Both retire on April 30, 2027, and Mistral lists Mistral Large 4 as the alternative for each.
What changed
mistral-medium-3-5 (Mistral Medium 3.5, version 26.04): deprecated October 31, 2026, retires April 30, 2027. Alternative: Mistral Large 4.
mistral-large-2512 (Mistral Large 3, version 25.12): same dates, same alternative.
Three audio models share the dates: voxtral-mini-tts-2603 (alternative Voxtral TTS 3), voxtral-mini-2602 (alternative Voxtral Mini Transcribe 3) and voxtral-mini-transcribe-realtime-2602 (alternative Voxtral Mini Transcribe Realtime 3).
Z.ai GLM 5.2 (zai-glm-5-2) retires on Mistral's platform today, October 31. Mistral's changelog names Z.ai GLM 5.3 as the replacement, at the same price.
Prices on RouterPlus
Prices per million tokens, from the live RouterPlus catalog on October 31, 2026:
Model
Provider
Input
Cached input
Output
Mistral Medium 3.5 (mistral-medium-2604)
Mistral AI
$1.50
$0.15
$7.50
Mistral Large 3 (mistral-large-2512)
Mistral AI
$0.50
$0.05
$1.50
Mistral Large 4 (mistral-large-4)
Mistral AI
$0.70
$0.07
$2.10
At these prices, moving from Medium 3.5 to Large 4 lowers the per-token cost. Moving from Large 3 to Large 4 raises it, by $0.20 per million input tokens and $0.60 per million output tokens. Large 4 is a different model, so rerun your evals before you switch.
Particle launched Particle-1 on October 28, 2026. It is a coding model post-trained on DeepSeek V4.1 Flash to solve tasks with fewer reasoning and tool-call tokens. It is available through Particle's API and in the RouterPlus catalog.
What changed
Particle-1 is DeepSeek V4.1 Flash post-trained with reinforcement learning, using a rank-64 LoRA adapter, according to Particle.
This first release focuses on software engineering. Particle says it plans to bring the approach to other domains in later releases.
Particle reports about 45% fewer tokens on average. On four coding benchmarks it reports 28.4% to 52.2% fewer output tokens and 34.4% to 51.6% lower average cost per task than DeepSeek V4.1 Flash, with pass rates within the base model's confidence intervals.
In the RouterPlus catalog the model ID is particle-ai/particle-1, with a 1,048,576-token context window, 131,072 maximum output tokens and one provider, Particle AI.
Prices on RouterPlus
Prices per million tokens, from the live RouterPlus catalog on October 31, 2026:
Model
Provider
Input
Cached input
Output
Particle-1
Particle AI
$0.2125
$0.0255
$0.85
DeepSeek V4.1 Flash 1017 (lowest listed)
Sail Research
$0.09
$0.006
$0.30
DeepSeek V4.1 Flash 1017
Particle AI
$0.2125
$0.0255
$0.85
Particle's launch post computes its cost-per-task figures at $0.30 input, $0.006 cached input and $1.20 output per million tokens for both Particle-1 and DeepSeek V4.1 Flash.
What it means
Particle's claim is about cost per task, not price per token. Savings depend on how many tokens your own workload burns, so compare the two models on your real traffic before switching. Both are callable behind one RouterPlus key.
What to do
curl https://api.routerplus.com/v1/chat/completions \
-H "Authorization: Bearer $ROUTERPLUS_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model": "particle-ai/particle-1", "messages": [{"role": "user", "content": "Fix the failing test in utils.py"}]}'
New SDK majors send the new version by default, so an upgrade can break working calls.
Anthropic released Claude API version 2026-10-21 on October 21, 2026, together with new SDK and CLI majors that send it by default. On that version, a request that sets temperature, top_p or top_k returns a 400 error on every model.
What changed
temperature, top_p and top_k are removed from the Messages API. Sending any of them returns a 400 error.
A document block placed directly in an assistant message returns a 400 error, as does a text block with a non-empty citations array in a user message.
These SDK versions send 2026-10-21 by default: Python 2.0.0, TypeScript 1.0.0, C# 13.0.0, Go 2.0.0, Java 3.0.0, PHP 1.0.0, Ruby 2.0.0 and the ant CLI 2.0.0. They also carry breaking changes unrelated to the API version and move from the MIT License to Apache 2.0.
Admin API responses rename fields, for example roles becomes role_ids and API keys return scope.workspace_id instead of workspace_id.
The version is available on the Claude API, Claude Platform on AWS, Claude Models on Google Cloud and Microsoft Foundry. On Amazon Bedrock it covers Claude 4.7 and later models.
What it means
An SDK upgrade is now a code change. Wrappers, retry helpers and framework defaults that send temperature on every call will start failing. Anthropic's versioning policy keeps existing input parameters within a given version, so calls pinned to 2023-06-01 are not affected by these removals.
What to do
# Before upgrading the Anthropic SDK, find sampling parameters
grep -rnE "temperature|top_p|top_k" src/
# Remove them from Messages and Message Batch requests,
# or stay on API version 2023-06-01 until you do.
Also in Anthropic's release notes: on October 30 it deprecated fast mode for Claude Opus 4.8, with removal on November 30, 2026. After that, claude-opus-4-8 requests with speed: "fast" return an error.
A 1.05M-token context, text and image input, and half GPT-5.6 Terra's input price.
OpenAI released GPT-6.1 Terra in its API as gpt-6.1-terra on October 26, 2026. Standard pricing is $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens.
What changed
Model ID gpt-6.1-terra, with text and image input and text output.
A 1,050,000-token context window, 922,000 maximum input tokens and 128,000 maximum output tokens. Knowledge cutoff is June 1, 2026.
Prompts over 272K input tokens are billed at 2x input and cache rates and 1.5x output for the whole request. Batch and Flex are 50% lower; regional processing adds 10%.
Tool calling requires the Responses API. Chat Completions and Batch are supported.
OpenAI says the model does not support UAE inference residency.
Pricing
Prices per million tokens. GPT-6.1 Terra is OpenAI's list price; the other rows are from the live RouterPlus catalog on October 31, 2026, provider OpenAI (direct API):
Model
Source
Input
Cached input
Output
gpt-6.1-terra
OpenAI list price
$1.00
$0.05
$5.00
gpt-5.6-terra
RouterPlus catalog
$2.00
$0.20
$12.00
gpt-6.1-sol
RouterPlus catalog
$2.00
$0.10
$10.00
gpt-6-luna
RouterPlus catalog
$0.10
$0.01
$0.50
gpt-6.1-terra is not in the RouterPlus catalog as of October 31, 2026.
What it means
At list price, GPT-6.1 Terra halves GPT-5.6 Terra's input price and cuts output by about 58%. Long prompts cost more: past 272K input tokens the whole request is billed at the higher rate, so check where your prompt sizes land.
Open weights under OpenMDW-1.1, built for recorded audio, hosted by Baseten, DeepInfra and Simplismart.
NVIDIA released Nemotron 3.5 Transcribe on October 28, 2026, an open-weight speech model that turns recorded audio into text in 40 languages. The weights are on Hugging Face and ModelScope under the OpenMDW-1.1 license.
What changed
The model is NVIDIA-Nemotron-3.5-Transcribe-30B-A4B. NVIDIA says it pairs a speech encoder and Nemotron 3 Diarization with the Nemotron 3.5 Lightning 30B-A3B language model, for about 4.3B active parameters.
One checkpoint transcribes 40 languages, translates 31 of them to or from English and attributes speakers in 13. It handles recordings up to one hour and labels up to eight speakers, with word-level timestamps on request.
The transcript style (verbatim, written form, speaker labels, translation) is set per request with a plain-language instruction.
It is built for recorded audio, not live streaming.
vLLM and SGLang-Omni support it from day one. Hosted endpoints are available from Baseten, DeepInfra and Simplismart. NVIDIA's announcement lists no prices.
What it means
Teams transcribing calls, meetings or dictation in batches now have a permissively licensed open model they can self-host or call through a host. Spot-check names and speaker changes against the recording, as NVIDIA advises, and compare cost per audio hour on your own files.
What to do
Download nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B from Hugging Face or ModelScope and serve it with vLLM or SGLang-Omni. The model card lists recommended prompt presets and which tasks each language supports.
Nemotron 3.5 Transcribe is not in the RouterPlus catalog. If you need dedicated capacity for it, tell us on I need inference.
RouterPlus lists eight of the newest open-weight models from as little as $0.039 per million input tokens. These are the lowest prices per million tokens for each, from the live catalog on October 31, 2026.
What changed
Qwen3.8 27B has the lowest input price in the table, $0.039 on Chutes.
GLM 5.3 Flash has the lowest output price, $0.20, and its lowest input price, $0.05, comes from the same provider, Wafer.
GLM 5.3 has the most providers in the catalog, 11.
Mistral released the Mistral Large 4 weights on October 29 and set its API price at $0.70 input and $2.10 output. RouterPlus lists it on 1 provider.
Reflection AI released the Beam weights on October 16 under Apache 2.0. RouterPlus lists it on 1 provider.
Lowest prices by model
Prices per million tokens. The lowest input and lowest output for a model can come from different providers.
DeepSeek released DeepSeek-V4.1-Flash-1017, an update to V4.1-Flash, on October 17, 2026. It is live on DeepSeek's website, app and API, and the weights are on Hugging Face and ModelScope.
What changed
deepseek-flash on DeepSeek's API is now served by DeepSeek-V4.1-Flash-1017. DeepSeek says the calling method and pricing are unchanged.
The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work. Their models are retired, so those requests go to V4.1-Flash-1017 at the Flash price.
DeepSeek lists a 1M-token context and a 384K maximum output.
The Hugging Face model card lists the MIT License for the weights.
DeepSeek says the update finishes tasks faster with shorter outputs. Its change log lists DeepSWE v1.1 at 77.0 (up from 74.2) and Terminal-Bench 4.0 at 42.3 (up from 31.2).
DeepSeek API prices
Per million tokens for deepseek-flash, from DeepSeek's pricing page:
Rate
Peak
Off-peak
Input, cache miss
$0.30
$0.15
Input, cache hit
$0.006
$0.003
Output
$1.20
$0.60
Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays. All other hours are off-peak.
Other hosts
DigitalOcean added DeepSeek V4.1 Flash 1017 to Serverless Inference, Dedicated Inference and Managed Agents at $0.30 input, $1.20 output and $0.006 cached input per million tokens.
Alibaba Cloud Model Studio added deepseek-v4.1-flash-1017 on October 17.
DeepSeek V4.1 Flash 1017 on RouterPlus
RouterPlus lists the model on 5 providers. Prices per million tokens, from the live catalog on October 31, 2026:
Provider
Input
Output
Sail Research
$0.09
$0.30
Wafer
$0.13
$0.50
RunInfra
$0.15
$0.60
Doubleword
$0.15
$0.60
Particle AI
$0.2125
$0.85
What to do
If you call the September V4.1 Flash on RouterPlus, moving to the update is a model ID change:
Web search and page fetch for agents, billed per successful request.
DeepSeek opened its Search and Fetch APIs to all developers on October 17, 2026. Each costs $0.75 per 1,000 successful requests. Calls made by server-side tools in DeepSeek Harness, Claude Code and Codex were free for DeepSeek API users, and DeepSeek set billing for them to start at 16:00 UTC on October 31, 2026 (00:00 Beijing time on November 1).
What changed
Search, at POST https://api.deepseek.com/v1/tools/search, returns structured web results from DeepSeek's in-house search engine.
Fetch, at POST https://api.deepseek.com/v1/tools/fetch, retrieves one web page and returns it as Markdown.
Both use the same bearer API key as the DeepSeek API.
Search returns 1 to 50 results per call (10 by default). The count setting does not change the price.
Server-side Search and Fetch calls from DeepSeek Harness, Claude Code and Codex move from free to paid at 16:00 UTC on October 31.
Pricing
From DeepSeek's pricing page and Search and Fetch guide:
API
Unit
Price
Search
1,000 successful requests
$0.75
Fetch
1,000 successful requests
$0.75
A request is billed only when it returns HTTP 200 with a result: a non-empty results array for Search, page content for Fetch. Failed requests, empty results and timeouts are not billed. These charges are separate from model token charges.
Rate limits
Per account, Search is limited to 100 requests per second and Fetch to 25. Fetch also has a per-domain limit. Requests over a limit get HTTP 429.
What to do
If your agent runs on DeepSeek's API through Claude Code, Codex or DeepSeek Harness, check how often it searches and fetches. From 16:00 UTC on October 31, those calls cost $0.75 per 1,000 each.
Search and Fetch are DeepSeek APIs, billed to a DeepSeek API key. For the model side, RouterPlus lists DeepSeek V4.1 Flash 1017 on 5 providers: DeepSeek V4.1 Flash 1017 API pricing.
A post-trained Kimi K3 with 1M context, plus a Highspeed variant.
Moonshot AI launched Kimi K3.1 Pro on October 23, 2026, together with a faster variant, Kimi K3.1 Pro Highspeed. Both are on the Kimi API now, and Moonshot says the full model weights will be released by November 4.
What changed
New model IDs on the Kimi API: kimi-k3.1-pro and kimi-k3.1-pro-highspeed.
Moonshot describes K3.1 Pro as a post-trained version of Kimi K3 with a 1M-token context window.
Moonshot says Highspeed runs at 2 to 3 times the speed of K3.1 Pro.
Thinking is always on. reasoning_effort takes low, high or max, and the default is max.
max_completion_tokens defaults to 131,072 and can be set up to 1,048,576.
Moonshot API prices
Per million tokens, from Moonshot's pricing page:
Model
Input
Output
Cached input
kimi-k3.1-pro
$2.00
$10.00
$0.20
kimi-k3.1-pro-highspeed
$4.00
$20.00
$0.20
kimi-k3
$3.00
$15.00
$0.30
Cache writes are billed separately. Batch requests for kimi-k3.1-pro and kimi-k3 cost 60% of the standard price.
On RouterPlus
Kimi K3.1 Pro is not in the RouterPlus catalog yet. Kimi K3 is available today on 9 providers. From the live catalog on October 31, 2026, the lowest input price is $1.50 (Sail Research) and the lowest output price is $9.50 (Chutes), per million tokens.
Mistral added Z.ai's GLM 5.3 Flash to its API on October 22, 2026, as zai-glm-5-3-flash. It costs the same on Mistral as on Z.ai's own API: $0.15 input, $0.03 cached input and $0.50 output per million tokens.
What changed
Mistral now serves GLM 5.3 Flash as zai-glm-5-3-flash, a multimodal mixture-of-experts model with 320B total and 18B active parameters.
Mistral lists a 1M-token context, a 128k output limit, and text and image input.
Mistral says the model is served without Mistral modifications.
The weights are on Hugging Face as zai-org/GLM-5.3-Flash under the MIT license.
Mistral retires GLM 5.2 (zai-glm-5-2) on its platform on October 31, 2026 and points users to GLM 5.3 (zai-glm-5-3) at the same price.
Z.ai released GLM 5.3 Flash on its own platform in August 2026. The October 22 change adds Mistral as a place to call it.
Pricing at the source
Prices per million tokens, from each vendor's pricing page:
Provider
Input
Cached input
Output
Mistral
$0.15
$0.03
$0.50
Z.ai
$0.15
$0.03
$0.50
GLM 5.3 Flash on RouterPlus today
RouterPlus lists GLM 5.3 Flash on 10 providers. Mistral is not one of them. Prices per million tokens, from the live catalog on October 31, 2026:
Calling GLM 5.2 on Mistral: Mistral's named replacement is zai-glm-5-3, at the same price. See GLM 5.3 API pricing.
Calling GLM 5.3 Flash through RouterPlus: use the model ID z-ai/glm-5.3-flash with one API key at https://api.routerplus.com/v1. Sign up at app.routerplus.com/login and call the API right away.
816B total, 42B active, 1M context, weights under the Muse Model License.
Meta released the weights for Muse Spark Open on October 15, 2026. It is an open-weight sparse mixture-of-experts model with 816B total and 42B active parameters and a 1M-token context. Meta says it is built for coding and agent work.
What changed
Weights are on Hugging Face in two checkpoints: meta-models/Muse-Spark-Open-816B-A42B (BF16) and meta-models/Muse-Spark-Open-816B-A42B-NVFP4.
The license is the Muse Model License, which Meta describes as having limited commercial terms. Hugging Face downloads need an access request.
Input is text, image, video and audio. Output is text.
vLLM shipped day-0 support, with prebuilt images for Nvidia and AMD.
Meta names Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI as API partners.
Checkpoint sizes
Meta's figures for the weight files alone, before KV cache and runtime overhead:
Checkpoint
Weight footprint
BF16
About 1.65 TB
NVFP4
About 537 GB
Pricing
OpenRouter lists Muse Spark Open at $1.25 input and $4.25 output per million tokens, with cache reads at $0.15.
RouterPlus lists it on 2 providers. Prices per million tokens, from the live catalog on October 31, 2026:
Provider
Input
Cached input
Output
Nebius
$1.25
$0.15
$4.25
Doubleword
$1.25
$0.15
$4.25
What to do
On RouterPlus, the model ID is meta/muse-spark-open, called with one API key at https://api.routerplus.com/v1. Sign up at app.routerplus.com/login and call the API right away.
RouterPlus does not have a Muse Spark Open model page yet. Browse every model on the RouterPlus models page.
1M context, $0.15 input, and weights now on Hugging Face.
MiniMax made M3.1 Flash generally available on October 19, 2026, through the MiniMax API, its M Plan subscription and the MiniMax Code app. On MiniMax's API it costs $0.15 input and $0.60 output per million tokens, with cache reads at $0.015.
What changed
The API model ID is MiniMax-M3.1-Flash.
MiniMax lists a context window of up to 1M tokens, with text, image and video input and text output.
The weights are now public on Hugging Face as MiniMaxAI/MiniMax-M3.1-Flash. MiniMax's October 19 post said they would come later.
The model card describes a mixture-of-experts model with about 428B total and 23B active parameters.
MiniMax says an early version ran anonymously as Space Bunny.
MiniMax says overall usage cost on agent tasks is 70% lower than its previous model. That is MiniMax's own figure, not a unit-price comparison.
License
The weights use the MiniMax Community License. It allows non-commercial use free of charge. Commercial use requires a visible "Built with MiniMax M3.1-Flash" credit and a one-time notice to MiniMax. Products or services above $20 million in yearly revenue need prior written authorization from MiniMax.
Pricing
Prices per million tokens on MiniMax's API:
Item
Price
Input
$0.15
Cached input
$0.015
Output
$0.60
RouterPlus lists MiniMax M3.1 Flash on 1 provider. Prices per million tokens, from the live catalog on October 31, 2026:
Provider
Input
Cached input
Output
StreamLake
$0.15
$0.015
$0.60
What to do
On RouterPlus, the model ID is minimax/minimax-m3.1-flash, called with one API key at https://api.routerplus.com/v1. Sign up at app.routerplus.com/login and call the API right away.
RouterPlus does not have a MiniMax M3.1 Flash model page yet. Browse every model on the RouterPlus models page.
Weights are on Hugging Face. API price drops to $0.70 input, $2.10 output.
Mistral released the open weights of Mistral Large 4 on Hugging Face on October 29, 2026, three weeks after the model entered public preview on October 6. The same day, Mistral cut the API price to $0.70 input and $2.10 output per million tokens.
What changed
The weights are on Hugging Face as mistralai/Mistral-Large-4-1T-A52B, in FP8 and NVFP4 checkpoints.
The license is the Mistral Ecosystem Open-Weight License, a modified MIT license. Businesses that run a Model-as-a-Service offering need a separate commercial license from Mistral.
The model card lists about 1T total parameters, 49B active per token (52B counting embeddings and output layers), a 1M-token context, and text and image input with text output.
Reasoning effort now has three levels: low, medium and high, with high the default.
Mistral says the released model had three more weeks of reinforcement learning than the preview. It reports 41% on Terminal-Bench 4.0, 13 points above the preview.
Mistral API pricing
Mistral's list price per million tokens for mistral-large-4:
Period
Input
Cached input
Output
Preview list price
$1.36
$0.14
$4.18
From October 29, 2026
$0.70
$0.07
$2.10
Mistral Large 4 on RouterPlus
RouterPlus lists mistral-large-4 on 1 provider. Prices per million tokens, from the live catalog on October 31, 2026:
Provider
Input
Cached input
Output
Mistral AI
$0.70
$0.07
$2.10
What to do
To call it through RouterPlus, sign up at app.routerplus.com/login and use the model ID mistral-large-4 with the base URL https://api.routerplus.com/v1.
To self-host, download the weights from Hugging Face and read the license first if you plan to offer the model as a hosted service.
501B total, 23B active, 1M context. On RouterPlus from $0.25 input.
Reflection AI released the weights of Beam, its first open-weight model, on Hugging Face on October 16, 2026, under the Apache 2.0 license. Reflection first announced Beam on October 5.
What changed
Beam is a sparse mixture-of-experts model with 501B total and 23B active parameters. The model card lists 501.3B and 23.4B.
It takes text in and returns text, with a 1M-token context and a knowledge cutoff of June 30, 2026.
Reflection says it pretrained Beam on 23.8 trillion tokens.
Weights ship in BF16, FP8 and NVFP4, plus a speculative decoding drafter. The main repository is reflection/Beam-1.0-501B-A23B.
Reflection's own API at platform.reflection.ai is a public beta that Reflection says is free to start.
Benchmarks reported by Reflection
Scores for the BF16 checkpoint, as reported in Reflection's model card:
Benchmark
Beam (BF16)
SWE-bench Verified
81.6%
Terminal Bench v2.1
80.1%
GPQA Diamond
90.6%
MCP Atlas
80.8%
Beam on RouterPlus
RouterPlus lists reflection/beam-1.0-501b-a23b on 1 provider. Prices per million tokens, from the live catalog on October 31, 2026:
Provider
Input
Cached input
Output
Nebius
$0.25
$0.05
$1.00
What to do
To call it through RouterPlus, sign up at app.routerplus.com/login and use the model ID reflection/beam-1.0-501b-a23b with the base URL https://api.routerplus.com/v1.
To self-host, the weights are on Hugging Face under Apache 2.0.
Moonshot AI stopped selling its Kimi K2 models to new accounts on its own API on October 31, 2026. Existing Moonshot customers keep access for now, and Moonshot plans to retire every K2 model by the end of December 2026.
What changed
New Moonshot API accounts can no longer use kimi-k2.6, kimi-k2.7-code or kimi-k2.7-code-highspeed.
Existing Moonshot accounts are not affected yet.
Moonshot plans to retire all K2 models by the end of December 2026.
kimi-k2.5 was already discontinued on Moonshot's API on August 31, 2026.
The notice covers Moonshot's own API. It does not say what third-party hosts will do.
Kimi K2.6 on RouterPlus today
RouterPlus lists Kimi K2.6 on 4 providers. Prices per million tokens, from the live catalog on October 31, 2026:
Kimi K3 is on 9 providers in the RouterPlus catalog. The lowest input price is $1.50 (Sail Research) and the lowest output price is $9.50 (Chutes). On RouterPlus, switching is a model ID change:
