SIENGSUPER INTELLIGENCE ENGINEERING

A practical guide to the models available now

What's new.
What it costs.
What you can build.

Current proprietary and open-weight models, with source-checked availability, licensing and workload costs.

Explore current models ↓Estimate a workload →
Sources checked · More than 24 hours old; verify details with providers.Source-linked · No synthetic scores
01 / Current catalog

The current model landscape

Newest verified releases first. Preview access and commercial-use restrictions are explicit.

Downloadable weights

Qwen Image 2.1 Turbo

Qwen/Qwen-Image-2.1-Turbo

Alibaba Qwen · 2026-10-09

7B visual-generation checkpoint for text-to-image and image editing, using an 8-step denoising schedule.

Availability: Public, ungated weights are downloadable now; commercial use requires a separate license.

Qwen Research License Agreement
Research and evaluation only under the supplied license. Commercial use needs separate permission; redistribution, attribution and derivative-model naming conditions also apply.

Public weights first verified October 9. Repository created at 04:50:01 UTC; the exact time it became public is unverified. The license’s September 20 date is not this checkpoint’s release date.

Hosted API

Claude Haiku 5.5

claude-haiku-5-5

Anthropic · 2026-10-07

Small, latency-focused model for summarization, classification, compaction and narrowly scoped agent work.

Availability: Available through the Claude API and supported AWS, Google Cloud and Microsoft platforms.

Public/API release; official Claude Platform release notes.

Hosted API

Mistral Large 4

mistral-large-4

Mistral AI · 2026-10-06

Natively multimodal mixture-of-experts preview for reasoning, coding and agent workflows.

Availability: Public preview API on Mistral Studio. Weights are announced for the end of October and are not verified available today.

Public API preview launch. A planned weights release is a separate future event.

Restricted access

Gemini 4 Argon

Gemini 4 Argon

Google DeepMind · 2026-09-30

Frontier model announced for complex software engineering, enterprise knowledge work and cybersecurity defense.

Availability: Limited to trusted cyber defenders through the Fairwind Program. No general-public API model identifier verified; the ID shown here is the announcement name.

Announcement and restricted rollout; this is not a general-public release date.

Hosted API

GPT-6.1 Sol

gpt-6.1-sol

OpenAI · 2026-09-29

General-purpose reasoning model for coding, computer use and professional work.

Availability: Available through the OpenAI API, ChatGPT Work and Codex on supported plans.

API release; official OpenAI changelog. October 8 added Ultrafast mode, not a new model release.

Hosted API

Claude Sonnet 5.5

claude-sonnet-5-5

Anthropic · 2026-09-28

Model for everyday professional tasks, bug fixes and document, slide and spreadsheet creation.

Availability: Available through the Claude API and supported AWS, Google Cloud and Microsoft platforms.

Public/API release; official announcement and Claude Platform release notes.

Hosted API

GPT-6 Luna

gpt-6-luna

OpenAI · 2026-09-22

Efficient text-and-image reasoning model for focused, high-volume workloads.

Availability: Available through Responses and Chat Completions; Decisions API support is in beta.

API release; official OpenAI changelog. October 6 added Decisions API beta support.

Hosted API

Claude Opus 5.5

claude-opus-5-5

Anthropic · 2026-09-22

Model for demanding coding and knowledge work, with adaptive reasoning and a 1M-token context window.

Availability: Available through the Claude API and supported cloud platforms.

Public/API release; official announcement and Claude Platform release notes.

Hosted API

Grok 4.7

grok-4.7

SpaceXAI / xAI · 2026-09-21

Text-and-image reasoning model for coding, tool-using agents and knowledge work.

Availability: Available through the xAI API as grok-4.7.

Public release; official launch announcement.

Downloadable weights

DeepSeek V4.1 Flash

deepseek-ai/DeepSeek-V4.1-Flash

DeepSeek · 2026-09-10

Multimodal mixture-of-experts model with native visual understanding and an asymmetric encoder-decoder architecture.

Availability: Public model weights; first-party API identifier is deepseek-flash. Older V4 API aliases route to this model.

MIT
Permissive license for the repository and weights; retain the required copyright and permission notice.

Public/API release; official DeepSeek release note.

Hosted API

GPT-6 Astra

gpt-6-astra

OpenAI · 2026-09-03

OpenAI’s high-capability model for complex, end-to-end work, including research, coding and document creation.

Availability: Available through the OpenAI API; product and plan availability varies.

API release; official OpenAI changelog.

Hosted API

Gemini 3.8 Flash

gemini-3.8-flash

Google DeepMind · 2026-09-02

Multimodal workhorse for reasoning and agent workflows; accepts text, images, video, audio and PDFs and returns text.

Availability: Stable Gemini API model; available through Google AI Studio and supported Google products.

Public/API release; official Google announcement.

Hosted API

Claude Fable 5.1

claude-fable-5-1

Anthropic · 2026-09-01

Model for ambitious, long-running coding, research and knowledge-work projects.

Availability: Generally available through Claude and the Claude API, with model-specific safeguards and data-retention requirements.

Public/API release; official Claude Platform release notes.

Downloadable weights

Qwen3.8 Flash-Next

Qwen/Qwen3.8-Flash-Next

Alibaba Qwen · 2026-08-26

Multimodal architecture preview with sparse mixture-of-experts, hybrid attention and n-gram embeddings.

Availability: Public model weights. Qwen3.8-Flash is a separately named managed production variant with additional serving features.

Qwen Community License 1.0
Commercial Model-as-a-Service or AI Work Assistant businesses need a separate license, except qualifying internal use. Large products also face model-name display requirements. See the license for definitions and thresholds.

Model-weight publication date supported by the official repository’s initial upload. Repository creation predates this; neither timestamp proves the exact public-release hour.

Downloadable weights

GLM-5.3 Flash

zai-org/GLM-5.3-Flash

Z.ai · 2026-08-26

Native multimodal mixture-of-experts model with 320B total and 18B active parameters, designed for coding and agent workflows.

Availability: Public model weights and first-party API access as glm-5.3-flash.

MIT
Permissive commercial-use license; retain the required copyright and permission notice.

Public release; dated official Z.ai release notes and repository weight uploads.

Downloadable weights

Qwen3.8 27B

Qwen/Qwen3.8-27B

Alibaba Qwen · 2026-08-14

27B-parameter vision-language model for reasoning, coding and multimodal agent tasks.

Availability: Public model weights in Hugging Face Transformers format.

Apache 2.0
Permissive commercial-use license with notice, attribution and other Apache 2.0 conditions.

Public weights release; dated news entry in Qwen’s official Qwen3.8 repository.

Downloadable weights

GLM-5.3

zai-org/GLM-5.3

Z.ai · 2026-08-14

Coding and long-horizon reasoning model improved through post-training over the GLM-5.2 base.

Availability: Public model weights and first-party API access as glm-5.3.

GLM-5.3 License
Commercial use is permitted subject to conditions. Model-as-a-Service operators whose combined affiliate revenue exceeds US$10 billion over 12 months need Z.ai security review before commercial use.

Hosted-model announcement date in the official launch-page source. The API release notes are dated August 18; public weights followed in late August.

Downloadable weights

Kimi K3

moonshotai/Kimi-K3

Moonshot AI · 2026-07-16

2.8T-parameter native multimodal model for long-horizon coding, knowledge work and reasoning, with a 1M-token context window.

Availability: Public model weights; first-party API identifier is kimi-k3.

Kimi K3 License
Model-as-a-Service operators above US$20 million aggregate annual affiliate revenue need a separate agreement. Large products face branding requirements; qualifying internal use and official/certified hosted access are exempt from those two clauses.

Hosted-product/API launch in Kimi’s official product history. Public weights followed July 27.

02 / Workload costs

Put a price on your workload

Compare current first-party API rates. Estimates include applicable context tiers; a low token price is not a quality score.

Include reasoning tokens in billed output. Excludes caching, tools, media, retries, taxes and regional premiums. Provider limits and time-limited offers apply.

Standard uncached text prices · checked 2026-10-09 · USD per million tokens
ModelInputBilled output
GPT-6.1 Solgpt-6.1-sol$2.00$10.00
GPT-6 Lunagpt-6-luna$0.10$0.50
Claude Haiku 5.5claude-haiku-5-5$0.10$0.50
Claude Sonnet 5.5claude-sonnet-5-5$2.00$10.00
Claude Opus 5.5claude-opus-5-5$4.00$20.00
Claude Fable 5.1claude-fable-5-1$10.00$50.00
Gemini 3.8 Flashgemini-3.8-flash$0.75$3.75
Grok 4.7grok-4.7$2.00$6.00
DeepSeek V4.1 Flashdeepseek-flash$0.30$1.20
Mistral Large 4 (public preview)mistral-large-4$0.68$2.09
  • GPT-6.1 Sol. Standard global pricing. Above 272,000 input tokens, the full request costs 2x input and 1.5x output. Fast is 2x Standard, Ultrafast 6x; Batch/Flex are half-price. Regional processing adds 10%. These alternative modes and premiums are excluded.
  • GPT-6 Luna. Standard global pricing. Above 272,000 input tokens, the full request costs 2x input and 1.5x output. Fast is 2x applicable rates; Batch/Flex are half-price. Regional processing adds 10%. These alternative modes and premiums are excluded.
  • Claude Haiku 5.5. Above 100,000 total prompt tokens, the whole request uses the higher rates. Global routing; US-only inference adds 10%. Batch discounts excluded.
  • Claude Sonnet 5.5. No long-context surcharge. Global routing; US-only inference adds 10%. Standard output cap shown; the separate Batch API beta permits 300K output. Batch pricing excluded.
  • Claude Opus 5.5. No long-context surcharge. Global routing; US-only inference adds 10%. Fast mode costs $8/$40; Batch and Fast pricing excluded. Standard output cap shown.
  • Claude Fable 5.1. No long-context surcharge. Global routing; US-only inference adds 10%. Standard online output cap shown. Batch discounts excluded.
  • Gemini 3.8 Flash. Paid Standard tier. Output includes thinking tokens. Official model page gives separate input and output limits, not a shared total-context limit. No long-context price tier is listed. Batch/Flex/Priority, grounding and cache storage excluded. Introductory rates through December 31, 2026; $1.50 input/$7.50 output starting January 1, 2027.
  • Grok 4.7. Official guide states no separate text-output limit; the shared 500,000-token context still applies. At 200,000 input tokens or more, all request tokens use long-context rates. US regional endpoint adds 10%. Grok 4.7 Fast is not offered through the public API.
  • DeepSeek V4.1 Flash. Peak, uncached rates shown. API model ID deepseek-flash serves DeepSeek-V4.1-Flash. Peak windows are Monday-Friday 01:00-04:00 and 06:00-10:00 UTC, excluding Chinese public holidays. All other hours, weekends and Chinese public holidays are half-price. No separate long-context tier is listed.
  • Mistral Large 4 (public preview). Public preview released October 6, 2026. Official model card advertises 1M context, without an exact integer or separate output cap. Do not invent an output cap or treat the nominal context as a verified exact API ceiling. No separate long-context tier is shown. Launch prices are 50% off for two weeks from the October 6 release. Listed regular rates are $1.36 input/$4.18 output. Exact expiry time is not published; recheck before October 20.

Price sources, limits and verification dates

What about self-hosted open weights?

Downloadable weights do not make inference free. Enter your own infrastructure quote for a scenario estimate.

Enter a GPU-hour price to estimate monthly infrastructure cost.

Hardware fit, throughput and commercial license rights must be evaluated separately. This is your scenario, not a measured serving benchmark.

03 / Evidence

Compare the test, not just the score

Use the same task, model version, reasoning budget and evaluation setup. We don't assign scores to new releases from unrelated historical runs.

Independent model evaluation

Current benchmark results

Explore the evaluator's current model index, task-level results and cost methodology.

View Artificial Analysis →
Reproducible evaluation

Build your own test set

Evaluate task success, latency, retries and cost together. A routing combination needs its own measured outcome.

Read our comparison guide →
04 / News & research

Engineering developments

Primary-source facts, practical implications and explicit limits.

Security engineering · · Anthropic

AI vulnerability scanning meets the maintainer bottleneck

What happenedAnthropic announced an opt-in service offering eligible open-source projects recurring vulnerability scans. Reports are model-generated and can include a reproducer and candidate patch. Its selected early validation sample included 97 high- or critical-rated findings: 85 met its disclosure bar, 11 duplicated known findings, and one was invalid.

Engineering perspective · analysisFor engineering teams, the useful unit is a validated fix, not a finding. Budget for reproduction, deduplication, threat-model review, patch tests, and maintainer time before increasing scan volume.

Limits of the evidenceThese are provider-reported results from a selected sample, not a general false-positive rate. Anthropic notes inflated severity and misunderstood threat models. Candidate fixes still need validation.

Read the primary source
Source checked 2026-10-09
Evaluation · · OpenAI Alignment Research

LASER targets rare safety failures with active learning

What happenedOpenAI describes LASER, a pipeline combining embedding-based classifiers, uncertainty sampling, reasoning-model labeling, and diversity selection. It reports much lower grading compute than random sampling for finding comparable numbers of rare disallowed examples.

Engineering perspective · analysisUse targeted sampling to build challenging evaluation sets economically, but keep a separate representative sample when estimating production failure rates. Validate model-generated labels before using them as ground truth.

Limits of the evidenceThe efficiency result applies to a specific rare-event sampling objective. It is not a universal inference-cost improvement. A boundary-focused test set does not measure real-world failure prevalence.

Read the primary source
Source checked 2026-10-09
Training & safety · · OpenAI

From safety claims to testable engineering controls

What happenedOpenAI proposes safety-case guidance for frontier reinforcement-learning training. It covers alignment, containment, monitoring, approval, and incident investigation, with recommendations such as held-out incident backtests, immutable transcripts, response commitments, and fail-closed controls.

Engineering perspective · analysisTurn a broad assurance into a reviewable engineering argument: name the risk, define evidence, assign an owner, test pause controls, and trace downstream dependencies before release.

Limits of the evidenceThis is proposed guidance with implementation still underway, not proof that every control is operational or an industry standard. Its training scope does not settle deployment safety.

Read the primary source
Source checked 2026-10-09