New model launches, official specifications and independent data · Page guide updated October 1, 2026; individual release reviews are shown below
The AI Model Release Tracker follows important GPT, Claude, Gemini and Grok launches without turning every announcement into a winner. It puts first-party release facts beside independent benchmark availability so you can decide what is ready to test, what it costs and what is still unknown.
The short answer
- Use the release date and access status to check whether a new model should actually be visible in your account.
- Use the provider specifications to compare context, output limits, caching and headline API rates.
- Use the independent-data panel only when a matching benchmark row exists. “Pending” means HiseHub will not invent a score from launch material.
- Open the source before changing a subscription or production API because rollouts and documentation can change quickly.
New Google API model? Check Argon migration, output budgets and scheduled prices before changing a working integration.
HiseHub Release Watch
Track new AI models without confusing launch claims with benchmark evidence
See when a model launched, what its provider officially lists for price, context and access, and whether a comparable independent benchmark row has reached HiseHub yet.
12 releases shown
Gemini 4 Argon
API model: gemini-4-argon
Reasoning and agent-work model with a shared input/output budget and optional long decode. Migration also changes caching and response handling; a model-string replacement alone is not a compatibility test.
- Input / output
- Through Nov 30, 2026: $2 input / $10 output
From Dec 1, 2026: $4 input / $20 output
USD per 1M tokens, paid Standard through Nov 30, 2026; from Dec 1: $4 input / $20 output. - Cached input
- $0.10implicit cache hits only; $0.20 per 1M from Dec 1. No explicit cache support.
- Context
- 1.05M shared1,048,576 tokens shared by input, thinking and output; not 1M input plus 1M output.
- Max output
- 65,536 defaultLong decode can use the remaining shared window with background execution and an explicitly higher output limit.
Stable API model, not available on the API Free tier. A Gemini consumer subscription does not establish API access or credit.
- Intelligence
- 52.6
- Coding
- —
- Agentic
- —
- Output speed
- 79.7 tok/s
Highest-scoring available row: Gemini 4 Argon (High). Scores are a snapshot, not a universal winner.
Check the full pricing table ↗
Read Google’s migration and long-decode guide ↗
Argon documentation and scheduled Standard rates checked October 27, 2026; release date remains October 13. Other cards keep their earlier review scope. Independent scores, if available, use the separately dated cached feed; no HiseHub model/account test.
Claude Haiku 5.5
API model: claude-haiku-5-5
Adaptive thinking with Medium default effort. Recount older Haiku prompts and check response block types, manual thinking budgets and sampling parameters before migrating; no drop-in compatibility or fixed cost-saving percentage is assumed.
- Input / output
- Up to 100K prompt: $0.10 input / $0.50 output
Above 100K prompt: $0.50 input / $2.50 output
USD per 1M tokens; Claude API Standard. The prompt length selects the request rate tier. - Cached input
- $0.01 / $0.05read short / long; 5-minute writes $0.125 / $0.625; 1-hour writes $0.20 / $1 per 1M
- Context
- 1M1,000,000-token context window; longer context is not always the same price
- Max output
- 128Kstandard maximum; Batch beta differs
First-party API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS are documented. Verify the provider, account and billing route; this API launch is not a promise of consumer-plan access.
- Intelligence
- 43.4
- Coding
- —
- Agentic
- —
- Output speed
- 234.4 tok/s
Highest-scoring available row: Claude Haiku 5.5 (Max). Scores are a snapshot, not a universal winner.
Check the full pricing table ↗
Check the official launch availability ↗
Official API release, rate tiers, limits and migration boundaries checked October 8, 2026. Benchmark signals, when present, remain the separately dated independent cached feed; no HiseHub model/account test.
GPT-6.1 Sol
API model: gpt-6.1-sol
DevDay release for coding and delegated work. Ultrafast now supports Sol: available to API users with separate billing and rate limits, but Work / Codex access requires Pro $500 or an eligible Enterprise / Edu plan. Other self-serve plans do not unlock it by buying credits.
- Input / output
- $2 / $10USD per 1M tokens; Standard, up to 272K input
- Cached input
- $0.10cache read; 5% of input rate; writes $2.50; long-context rates differ
- Context
- 1.05M1,050,000 total; maximum input 922,000 tokens
- Max output
- 128K128,000 output tokens
Plus, Pro, Business, Enterprise and Edu in Work / Codex, subject to rollout; Enterprise and Edu require admin enablement. Free / Go are not included at launch. Not ordinary Chat.
- Intelligence
- 51.8
- Coding
- —
- Agentic
- —
- Output speed
- 73.5 tok/s
Highest-scoring available row: GPT-6.1 Sol (Max). Scores are a snapshot, not a universal winner.
Check the full pricing table ↗
Check Ultrafast plan and billing requirements ↗
Ultrafast access and billing boundaries rechecked October 11, 2026; the displayed numeric prices remain Standard API rates, not Ultrafast or subscription-credit prices. Launch facts retain their October 1 review scope. No HiseHub account or model-output test.
Claude Sonnet 5.5
API model: claude-sonnet-5-5
API High effort by default; app / Claude Code Medium by default. Base input/output rates match Sonnet 5, but thinking and tool integrations require migration checks. Original release facts reviewed October 1, 2026.
- Input / output
- $2 / $10USD per 1M tokens; Claude API Standard
- Cached input
- $0.10cache read; lowered October 7; 5-minute writes $2.50; 1-hour writes $4 per 1M tokens
- Context
- 1M1,000,000-token context window
- Max output
- 128Kstandard maximum output; Batch beta differs
Claude and Claude Code have plan limits; API/cloud billing is separate. Eligible Max and Team subscribers can claim first-party API credits, not interactive Claude Code or app extra usage. Self-run claude -p with a linked-organization API key is a different, covered Agent SDK route.
- Intelligence
- 56.0
- Coding
- —
- Agentic
- —
- Output speed
- 141.3 tok/s
Highest-scoring available row: Claude Sonnet 5.5 (Max, Default Fallback). Scores are a snapshot, not a universal winner.
Check the full pricing table ↗
Check subscriber API credit eligibility ↗
Only cache pricing and subscriber API-credit eligibility, including the self-run SDK versus interactive Code distinction, rechecked October 8, 2026. This is not a new model release or a new benchmark test.
Claude Opus 5.5
API model: claude-opus-5-5
New Opus release for sustained coding and knowledge work, with lower token rates and always-on adaptive thinking. Older tool integrations need migration checks. Official facts reviewed September 23, 2026.
- Input / output
- $4 / $20USD per 1M tokens; Standard processing
- Cached input
- $0.20cache read; 5-minute writes $5 per 1M tokens
- Context
- 1M1,000,000-token context window
- Max output
- 128Kstandard maximum output; Batch beta differs
Anthropic lists Pro, Max, Team and Enterprise access on Claude. API/cloud billing is separate. Account, workspace and usage conditions still apply.
- Intelligence
- 57.6
- Coding
- —
- Agentic
- —
- Output speed
- 98.3 tok/s
Highest-scoring available row: Claude Opus 5.5 (Max, Default Fallback). Scores are a snapshot, not a universal winner.
GPT-6 Sol
API model: gpt-6-sol
OpenAI positions Sol for complex coding and agentic work. API text/image input and text output; use Responses for built-in tools. Official facts reviewed September 23, 2026.
- Input / output
- $2 / $10USD per 1M tokens; Standard, up to 272K input
- Cached input
- $0.20cache read; writes 1.25x input; long-context rates differ
- Context
- 1.05M1,050,000 total; maximum input 922,000 tokens
- Max output
- 128K128,000 output tokens
Available in Work and Codex on eligible paid plans, subject to client, workspace and rollout. Not available in ordinary Chat. API usage has separate billing.
- Intelligence
- 47.6
- Coding
- —
- Agentic
- —
- Output speed
- 124.8 tok/s
Highest-scoring available row: GPT-6 Sol (Max). Scores are a snapshot, not a universal winner.
GPT-6 Luna
API model: gpt-6-luna
Lower-cost model for focused, high-volume tasks. Base input and output rates are one twentieth of Sol rates; actual workload costs depend on usage and accepted results. Official facts reviewed September 23, 2026.
- Input / output
- $0.10 / $0.50USD per 1M tokens; Standard, up to 272K input
- Cached input
- $0.01cache read; writes 1.25x input; long-context rates differ
- Context
- 1.05M1,050,000 total; maximum input 922,000 tokens
- Max output
- 128K128,000 output tokens
Free and Go list Standard-speed desktop access subject to rollout; eligible paid plans also offer Luna. Not ordinary Chat. API availability follows the project and key.
- Intelligence
- 38.1
- Coding
- —
- Agentic
- —
- Output speed
- 124.3 tok/s
Highest-scoring available row: GPT-6 Luna (Max). Scores are a snapshot, not a universal winner.
Grok 4.7
API model: grok-4.7
Released September 21 for coding, agentic tasks, and knowledge work. xAI lists configurable reasoning; HiseHub keeps provider claims separate from its independent benchmark feed.
- Input / output
- Short: $2 input / $6 output
Long: $4 input / $12 output
USD per 1M tokens; global Standard. Long rates apply to the whole request at ≥200K prompt tokens. - Cached input
- $0.50 short / $1 longUSD per 1M cached input tokens; cached input counts toward the prompt threshold.
- Context
- 500K500,000-token context window
- Max output
- See current API docsProvider does not list a separate cap on this page
Available on the xAI API as grok-4.7. Fast is listed for Cursor and Grok Build, not the public API; consumer Grok plan access is a separate check.
- Intelligence
- 46.4
- Coding
- —
- Agentic
- —
- Output speed
- 91.9 tok/s
Highest-scoring available row: Grok 4.7 (Xhigh). Scores are a snapshot, not a universal winner.
Check the full pricing table ↗
Grok card rechecked October 7, 2026. The US route adds 10%; exactly 200K follows the main pricing table, whose boundary wording differs from abbreviated model prose. Batch is not supported. No account or model-output test.
GPT-6 Astra
API model: gpt-6-astra
Adds asynchronous tool calls, mid-turn steering and effort changes for long-running OpenAI Responses API workflows.
- Input / output
- $10 / $50base; >272K input changes rates
- Cached input
- $1per 1M tokens
- Context
- 1.05M1,050,000 context window; max input 922,000 tokens
- Max output
- 128K128,000 output tokens
Current API documentation lists the model; verify your project entitlement and rate limits. ChatGPT plan access is a separate product check. Reviewed October 2, 2026.
- Intelligence
- 52.7
- Coding
- 76.9
- Agentic
- 51.0
- Output speed
- 65.6 tok/s
Highest-scoring available row: GPT-6 Astra (Max). Scores are a snapshot, not a universal winner.
Gemini 3.8 Flash
API model: gemini-3.8-flash
Targets long-horizon software, agent and enterprise workflows at a substantially lower token-price tier.
- Input / output
- $0.75 / $3.75introductory through Dec 31, 2026
- Cached input
- Supportedcheck the current pricing page
- Context
- 1.05M1,048,576 input tokens
- Max output
- 65.5K65,536 output tokens
Stable in the Gemini API. Gemini Apps access and plan quotas are a separate check; Google’s current help page lists the Gemini 3 models across tiers, while product, account, and regional availability can still differ.
- Intelligence
- 40.9
- Coding
- 76.3
- Agentic
- 40.2
- Output speed
- 134.0 tok/s
Highest-scoring available row: Gemini 3.8 Flash (High). Scores are a snapshot, not a universal winner.
Claude Fable 5.1
API model: claude-fable-5-1
Adds adaptive reasoning controls for sustained agentic work, with migration changes for tool choice and thinking blocks.
- Input / output
- $10 / $50per 1M tokens
- Cached input
- $0.25per 1M tokens
- Context
- 1M1,000,000 input tokens
- Max output
- 128K128,000 output tokens
Documented for the Claude API and multiple cloud platforms; Mythos 5.1 remains a separate restricted route.
- Intelligence
- 53.4
- Coding
- 81.6
- Agentic
- 57.9
- Output speed
- 69.8 tok/s
Highest-scoring available row: Claude Fable 5.1 (Max, Default Fallback). Scores are a snapshot, not a universal winner.
Grok 4.6
API model: grok-4.6
A newer Grok API generation focused on reasoning and agent work with a 500K-token context window.
- Input / output
- Short: $2 input / $6 output
Long: $4 input / $12 output
USD per 1M tokens; global Standard. Long rates apply to the whole request at ≥200K prompt tokens. - Cached input
- $0.50 short / $1 longUSD per 1M cached input tokens; cached input counts toward the prompt threshold.
- Context
- 500K500,000-token context window; not a separate maximum-input guarantee
- Max output
- Not separately listedCheck the live model documentation
Available through the xAI API; consumer-product access and API billing should be verified separately.
- Intelligence
- 44.3
- Coding
- 76.8
- Agentic
- 53.0
- Output speed
- 76.5 tok/s
Highest-scoring available row: Grok 4.6 (High). Scores are a snapshot, not a universal winner.
Check the full pricing table ↗
Grok card rechecked October 7, 2026. The US route adds 10%; exactly 200K follows the main pricing table, whose boundary wording differs from abbreviated model prose. Batch is not supported. No account or model-output test.
Two evidence layers, kept separate
Provider facts—release date, token price, context, model ID and access—come from the linked first-party documentation and are checked editorially. Benchmark signals come from HiseHub’s cached Artificial Analysis feed and update when that existing daily sync receives a matching model row.
The benchmark panel may update automatically after the independent source adds or revises a model. The editorial release facts do not silently scrape provider pages; HiseHub rechecks those when providers change documentation. Last benchmark sync: October 27, 2026 06:37 UTC.
Gemini 4 Argon: check the integration, not just the model name
Argon review — October 27, 2026: Google released gemini-4-argon on October 13. Its API availability is not a promise about the model picker in a Gemini consumer plan. The useful upgrade question is whether your request, response parser and budget still work—not whether a launch headline calls it the best.
- Reserve output room inside the context window. Argon shares 1,048,576 tokens across input, thinking and output. Its normal output ceiling is 65,536; the larger long-decode allowance needs background execution and a higher explicit output limit. You cannot budget the full window once for input and again for output. See the official Argon specifications.
- Retest the old request and response path. Explicit
cached_contentreferences are unsupported. Sampling controls such astemperature,top_pandtop_kare ignored. Long decode can return severalmodel_outputsteps, so a parser that reads only one may discard part of a completed answer. Compare the assembled response with its completion status and usage before retrying a costly job. Google’s Argon migration guide documents these changes. - Price the same workload on both sides of December 1. Paid Standard rates through November 30 are $2 per million uncached input tokens and $10 per million output tokens, including billed thinking; they become $4 and $20 on December 1. For an illustrative request with 50,000 uncached input and 20,000 billable output tokens, the token subtotal is $0.30 now, $0.60 at the scheduled rates. This arithmetic excludes tool charges, taxes, retries and other service tiers. It is not a measured cost per successful task. Confirm the current Google rate card before funding a project.
- Check storage before using real customer material. New paid Generate Content projects store requests by default from October 26; older projects retain their prior setting. The Interactions endpoint has its own storage behavior, and disabling it can affect saved conversation state. Review the chosen project, endpoint and request-level override in Google’s logging documentation. Do not treat “paid API” as a no-storage guarantee.
HiseHub’s recommendation: Start with a small, non-sensitive example that exercises your actual tools and parser. Set a spend ceiling, inspect the returned usage and keep your previous working configuration available. A successful text-only request does not validate a document workflow or a long-running agent.
This is a first-party documentation review and an illustrative calculation, not an account, latency or model-output test. Other release cards retain their own dated evidence; a cached benchmark score is not a billing or compatibility guarantee.
Grok cards: check the request tier, not only the headline rate
Grok card review — October 7, 2026: Grok 4.6 and 4.7 now show short and long Standard global prices separately. The whole prompt, including cached input, determines the tier; a 500K context window does not promise short-context pricing. The US route adds a token premium. At exactly 200K, the main pricing table and abbreviated model prose use different boundary wording, so reconcile actual charges rather than assuming the lower tier. Read the Grok API billing guide for worked examples and route checks.
This was a documentation review, not a model, latency or account test. Launch dates are unchanged; other providers’ cards retain their separate review history. Voice and Imagine use different rate units and are not included in these text-token comparisons.
AI Updates Official-source events; automatic candidates are filtered, deduplicated and labelled. Last source run: October 27, 2026 07:07 UTC. AnythingLLM v1.19.0 is listed as a stable release. Source notes: The same kind of built-in engine that ships with AnythingLLM Desktop can now run inside the container: Lemonade, the fully open-source local inference server from AMD, running your models with llama.cpp.; Pass your GPU into the container and the engine is accelerated on whatever hardware you have.; When it's turned on, Built-in becomes the default LLM provider during onboarding, with three models recommended for your hardware. HiseHub links the official source; rollout and compatibility remain source-specific. Affects: Teams evaluating or operating AnythingLLM in a hosted or self-hosted workflow. Action: Read the official notes first, check compatibility and backups, and verify the relevant plan, region or deployment before changing production. n8n 2.44.7 is listed as a stable release. Source notes: core: Start sub-workflow error workflow when a pre-run check fails; Microsoft SharePoint Node: Fix the column mapper 403 and empty items in Map Automatically; Update certifi to 2026.10.19 in task runner HiseHub links the official source; rollout and compatibility remain source-specific. Affects: Teams evaluating or operating n8n in a hosted or self-hosted workflow. Action: Read the official notes first, check compatibility and backups, and verify the relevant plan, region or deployment before changing production. Open WebUI v0.12.4 is listed as a stable release. Source notes: 🎚️ Chat Controls saved per setting. A change in a chat's Controls, such as its system prompt or a parameter, now saves only that setting, so a copy of the chat open in another tab or on another device no longer wipes it with older settings.; ✍️ Note edits by text. The model can now edit a note by finding and replacing exact text instead of only by character offsets, picking one match or all of them and getting the positions and surrounding text back when a match is ambiguous.; 🛰️ MCP protocol 2026-07-28. MCP connections now also work with servers that speak the stateless 2026-07-28 protocol version, while older servers still connect through the earlier handshake. HiseHub links the official source; rollout and compatibility remain source-specific. Affects: Teams evaluating or operating Open WebUI in a hosted or self-hosted workflow. Action: Read the advisory, compare affected versions with your deployment, and prioritize a tested upgrade or mitigation. Ollama v0.40.11 is listed as a stable release. Source notes: ollama run` plays text or a chat model's reply aloud as the audio is generated, and apps can stream audio through the OpenAI-compatible `/v1/audio/speech` endpoint.; d3 brings vLLM Semantic Router's decision models to Ollama.; Give a model text, structured state, or an image, and ask it to classify, answer yes or no, or score options with probabilities. HiseHub links the official source; rollout and compatibility remain source-specific. Affects: Teams evaluating or operating Ollama in a hosted or self-hosted workflow. Action: Read the official notes first, check compatibility and backups, and verify the relevant plan, region or deployment before changing production. Ollama v0.40.10 is listed as a stable release. Source notes: Nemotron 3.5 Bolt E4B is NVIDIA’s smaller, Gemma 4-based vision-language model for fast local conversations, image understanding, and tool use.; It supports a 128K-token context window.; The model includes a DFlash draft for speculative decoding. HiseHub links the official source; rollout and compatibility remain source-specific. Affects: Teams evaluating or operating Ollama in a hosted or self-hosted workflow. Action: Read the official notes first, check compatibility and backups, and verify the relevant plan, region or deployment before changing production. Dify 1.18.0 is listed as a stable release. Source notes: Dify 1.18.0 brings Agents that route workflows, portable app packages, deeper visibility into nested workflow runs, and the opt-in Agent Knowledge Base experience.; !IMPORTANT] > Read the upgrade guide below before deploying. Upgrade the API, workers, Agent Backend, and Agent runtime images together.; Local Agent sandboxes require a non-empty authentication token, and Agent Backend uses a new run protocol. HiseHub links the official source; rollout and compatibility remain source-specific. Affects: Teams evaluating or operating Dify in a hosted or self-hosted workflow. Action: Read the official notes first, check compatibility and backups, and verify the relevant plan, region or deployment before changing production. Ollama v0.40.9 is listed as a stable release. Source notes: ollama launch` now passes cloud models' context lengths and capabilities to coding tools, fixing incorrect context and compaction settings. HiseHub links the official source; rollout and compatibility remain source-specific. Affects: Teams evaluating or operating Ollama in a hosted or self-hosted workflow. Action: Read the official notes first, check compatibility and backups, and verify the relevant plan, region or deployment before changing production. Langflow v1.13.0 is listed as a stable release. Source notes: feat admin : add team and user administration API and CLI; feat authz : add enterprise directory provenance and CLI; feat: add FunASR transcription component HiseHub links the official source; rollout and compatibility remain source-specific. Affects: Teams evaluating or operating Langflow in a hosted or self-hosted workflow. Action: Read the official notes first, check compatibility and backups, and verify the relevant plan, region or deployment before changing production. Langflow v1.12.8 is listed as a stable release. Source notes: fix: normalize exported flow code before execution HiseHub links the official source; rollout and compatibility remain source-specific. Affects: Teams evaluating or operating Langflow in a hosted or self-hosted workflow. Action: Read the official notes first, check compatibility and backups, and verify the relevant plan, region or deployment before changing production. Ollama v0.40.8 is listed as a stable release. Source notes: Flint brings Recursive's coding model to Ollama.; It is designed for the work coding agents do across a repository: understanding unfamiliar code, tracing bugs, implementing features, and adding tests.; Reasoning, multi-turn tool use, and a 262K-token context window support longer tasks across files and tool interactions. HiseHub links the official source; rollout and compatibility remain source-specific. Affects: Teams evaluating or operating Ollama in a hosted or self-hosted workflow. Action: Read the official notes first, check compatibility and backups, and verify the relevant plan, region or deployment before changing production. Langflow v1.12.7 is listed as a stable release. Source notes: fix security : validate OpenSearch connector URLs; fix security : scope original-file S3 uploads; fix deps : dependency upgrades for 1.12.7 HiseHub links the official source; rollout and compatibility remain source-specific. Affects: Teams evaluating or operating Langflow in a hosted or self-hosted workflow. Action: Read the official notes first, check compatibility and backups, and verify the relevant plan, region or deployment before changing production. Ollama v0.40.7 is listed as a stable release. Source notes: pplx-decider is a 27B decision model from Perplexity, fine-tuned from Qwen3.8-27B; Continuing an OpenAI-compatible Responses conversation no longer fails when switching between models with and without thinking, or when the history contains another provider's compaction items; Fixed streaming Responses requests where client tool calls e.g. Codex with web search could emit before open reasoning items, producing misordered/duplicate tool call events HiseHub links the official source; rollout and compatibility remain source-specific. Affects: Teams evaluating or operating Ollama in a hosted or self-hosted workflow. Action: Read the official notes first, check compatibility and backups, and verify the relevant plan, region or deployment before changing production. LobeHub v2.2.20 is listed as a stable release. Source notes: Topics, documents and device data load from local copies first, and server runs can use models that only exist on your device.; Widgets & dashboards — Home and project dashboards, an agent widget tool, chat previews, and new `lh widget` / `lh dashboard` commands.; LLM relay to your device — Server-side runs, preset tasks, model lists and Ollama downloads can now use models that are only available in your browser tab or desktop app. HiseHub links the official source; rollout and compatibility remain source-specific. Affects: Teams evaluating or operating LobeHub in a hosted or self-hosted workflow. Action: Read the official notes first, check compatibility and backups, and verify the relevant plan, region or deployment before changing production. Open WebUI v0.12.3 is listed as a stable release. Source notes: 🔄 General improvements. Various improvements were implemented across the application to enhance performance, stability, and security.; 🌐 Translation updates. Translations for Portuguese Brazil , Chinese Traditional , Arabic, Bengali, Hebrew, Georgian, Punjabi, Slovak, Serbian, Urdu, Uzbek and Vietnamese were enhanced and expanded.; ⏩ Continue Response with several models. Continue Response now continues the reply it was clicked under, where in a chat with several models it could extend another model's reply. HiseHub links the official source; rollout and compatibility remain source-specific. Affects: Teams evaluating or operating Open WebUI in a hosted or self-hosted workflow. Action: Read the official notes first, check compatibility and backups, and verify the relevant plan, region or deployment before changing production. Ollama v0.40.5 is listed as a stable release. Source notes: NVIDIA Nemotron 3.5 Super is a vision-language model for coding, reasoning, tool use, and document understanding.; It pairs a 120B-parameter hybrid Mamba-2/Transformer mixture-of-experts backbone 12B active per token with a C-RADIO vision encoder, and supports tool calling, an on/off thinking mode, and context up to 1M tokens.; A compact sibling, Nemotron 3.5 Bolt, distills Super down to 74B parameters 9B active tuned to fit a 24 GB NVIDIA GPU for everyday agentic tasks, with speculative decoding via its included drafter. HiseHub links the official source; rollout and compatibility remain source-specific. Affects: Teams evaluating or operating Ollama in a hosted or self-hosted workflow. Action: Read the official notes first, check compatibility and backups, and verify the relevant plan, region or deployment before changing production. These are short, source-linked notices rather than long news articles. Release status, compatibility and access can change; verify the official source before upgrading or changing a plan.
Curated changes worth checking
AnythingLLM v1.19.0
n8n 2.44.7
Open WebUI v0.12.4
Ollama v0.40.11
Ollama v0.40.10
v1.18.0 — Smarter Agent Routing, Portable Apps, Deeper Workflow Visibility
Ollama v0.40.9
Langflow v1.13.0
Langflow v1.12.8
Ollama v0.40.8
Langflow v1.12.7
Ollama v0.40.7
LobeHub v2.2.20
Open WebUI v0.12.3
Ollama v0.40.5
September 28–29: Sonnet 5.5 and GPT-6.1 Sol
Read Sonnet 5.5 vs Opus 5.5 for effort, cost and API migration, and the DevDay practical guide for GPT-6.1 Sol, Dots and access conditions. The cards keep official specifications separate from independently synchronized benchmark rows.
September 22: Opus 5.5, GPT-6 Sol and Luna
The latest additions cover three different price tiers and access routes. Read our Opus 5.5 vs GPT-6 Sol vs Luna comparison for practical cost examples, plan access and the limits of current independent scores.
What changed in the September 2026 model wave?
GPT-6 Astra, Gemini 3.8 Flash and Claude Fable 5.1 arrived within three days. That makes a normal “which model is best?” headline especially unreliable: the products target different price bands, appear through different access routes and do not receive comparable third-party evaluation data at the same time.
GPT-6 Astra and Claude Fable 5.1 share the same published base token rates, but their caching and long-context billing differ. Gemini 3.8 Flash is listed in a much lower API price tier. A useful comparison therefore needs to separate model capability from availability, integration risk and total workload cost.
For a detailed early analysis, read GPT-6 Astra vs Claude Fable 5.1: Price, Context, Access & Early Verdict. For an interactive ranking after independent data arrives, use the AI Model Comparison.
How HiseHub verifies the tracker
Provider facts come from first-party documentation
Release dates, API model IDs, context windows, maximum output, token prices and stated access routes are checked against documentation from OpenAI, Anthropic, Google and xAI. Each release card links directly to its provider page. HiseHub does not treat a social post, reseller page or copied pricing table as the source of record.
Benchmark data comes from a separate independent feed
Capability scores and measured output speed are matched against the cached Artificial Analysis data already used by HiseHub’s comparison tool. That feed refreshes separately. When several reasoning or effort variants exist, the release card displays the highest-scoring available row and names it explicitly rather than hiding the variant.
This separation prevents three common errors: copying a provider’s self-reported launch benchmark into an “independent” ranking, assigning an old model family’s score to a new version, and calling an untested model the winner because it was announced most recently.
How to use a new-model announcement before paying
- Verify live access. A release page can describe a phased rollout before the model reaches every plan, account, region, workspace or API project.
- Check the correct billing route. A consumer subscription and an API project are normally separate products. One does not automatically include the other.
- Compare the complete rate card. Input, output, cached input, cache writes, tools, long-context multipliers and regional processing can affect the final bill.
- Wait for comparable evidence when necessary. Provider system cards are valuable, but they are not a neutral head-to-head evaluation.
- Test representative work. Measure successful completion, retries, latency and total cost on tasks that resemble your real use case.
Use the AI API Cost Calculator to turn token rates into a workload estimate. If your question is which consumer tier exposes a model, check the AI Model Access Map instead.
Why a release tracker is different from an AI news feed
A broad AI news feed becomes stale quickly and usually adds little value beyond rewriting announcements. This tracker is narrower: it follows launches that change a purchase, access or integration decision. Every entry needs a first-party source, structured specifications and a clear independent-data status.
That also means HiseHub will not add every preview, rumour, fine-tune or renamed product. A model belongs here when users can meaningfully ask whether they can access it, what it costs or whether independent comparison data is available.
Frequently asked questions
Does the AI Model Release Tracker update automatically?
The independent benchmark panel reads HiseHub’s existing daily model-data cache and can update when a matching model row appears. Official product facts are reviewed editorially because provider pages can change structure or wording and should not be copied blindly.
Why is a new model marked “benchmark pending”?
It means the independent feed does not yet contain a comparable scored row. HiseHub keeps that gap visible instead of estimating a result from an older model or a provider claim.
Do higher benchmark scores always mean a better subscription?
No. A consumer plan can have different usage limits, model access, tools and regional availability. Benchmarks can help form a shortlist, but the right plan depends on the task, access route, reliability and total price.
Are the prices consumer subscription prices?
No. The release cards show published text API rates per million tokens. For consumer and App Store pricing, use HiseHub’s AI Subscription Price Comparison.
Page guide updated October 1, 2026; release cards retain their separate review dates. See each source for current availability. Model names, access, prices, limits and third-party scores can change. HiseHub is independent and is not affiliated with OpenAI, Anthropic, Google, xAI, X or Artificial Analysis.