Monthly roundup
August 2026 in AI models
A release-wave summary for readers who want the important changes without pretending every launch is a reason to switch.
The signal
25 releases plus 1 material status update
Z.ai, Tencent, InclusionAI, SpaceXAI, Qwen, DeepSeek, OpenAI, Google, Microsoft AI, NVIDIA, Anthropic, Meta, Mistral, Sakana AI account for 25 tracked release events. 21 catalogued events link to a detailed comparison profile. 9 launch guides are available for deeper trade-off coverage. There is no single winner because the constraints are different.
At a glance
- Release events
- 25
- Profile links
- 21
What shipped and changed
Availability updateGLM-5.3 open weights
Z.ai
The promised GLM-5.3 weights dropped on Hugging Face (753B total MoE, FP8) under Z.ai's custom glm-5.3 licence — not a standard open-source licence.
Date note: The exact GLM-5.3 licence requires security review for model-as-a-service operators whose licensee/affiliate combined revenue exceeds $10B in any consecutive 12 months; embedded end-user features and mere relays are excluded from that definition. Current weight availability is verified; this illustrative event's exact initial publication day remains unverified.
Hugging Face — zai-org/GLM-5.3
Limited previewTencent Hy4 Preview
Tencent
Open-weight (Apache 2.0) flagship preview: 770B total / 49B active MoE with 1M context and an MTP speculative-decoding layer.
Date note: Weights on Hugging Face (repo created Aug 27). No verified hosted token pricing yet.
Tencent Hy4 Preview announcement
Availability updateLing-3.0-flash-Fin announcement
InclusionAI
Ant Ling announced its finance-enhanced checkpoint and planned weights for the following week.
Date note: Announcement date is separate from public-weight availability. September 3 repository creation does not prove the exact weight-release day; weights are currently downloadable.
Official Ant Ling announcement- General availability
GLM-5.3-Flash
Z.ai
Natively multimodal open-weight release (MIT): 320B/18B MoE, 1M context, hybrid sparse/linear attention. Revealed as the model behind the anonymous 'Ox Alpha' listings.
Date note: Z.ai API $0.15/$0.50 per MTok, cached $0.03; third-party tariffs need independent checks. AA public-table capture October 1 reports 42 for GLM-5.3-Flash on Intelligence Index v4.3.2. This capture is not the release-day evaluation.
Z.ai GLM-5.3-Flash announcement 
Availability updateGrok 4.6 on Microsoft Foundry
SpaceXAI
Grok 4.6 entered public preview on Microsoft Foundry Models (long-horizon reasoning positioning).
Date note: Enterprise Azure path alongside the Aug 19 Bedrock availability.
xAI — Grok 4.6 on Microsoft Foundry
Limited previewSecondary receiptQwen3.8-Flash-Next
Qwen
Experimental preview of the Qwen 4 architecture: 125B total / 6B active MoE plus a 51B n-gram embedding table, 262K native context (1M extensible), text/image/video input.
Date note: Open weights under the qwen-community-1.0 license (more restrictive than MIT/Apache). Vendor-reported benchmarks only; hosted qwen3.8-flash is listed at $0.15/$0.47 per MTok on Model Studio. Trackers place the drop between Aug 26 and Aug 28.
The New Stack — Qwen3.8-Flash previews Qwen4 (secondary source)
Limited previewDeepSeek V4-Flash-Vision-Exp
DeepSeek
Experimental image-input variant of V4-Flash on the DeepSeek API.
Date note: Retired together with V4-Flash on 2026-09-10; image input now ships in V4.1-Flash.
DeepSeek API changelog- Availability update
Grok 4.6
SpaceXAI
Amazon Bedrock added Grok 4.6 (model id `xai.grok-4.6`).
Date note: Enterprise AWS path; first-party xAI list rates in this catalog remain $2/$6.
AWS what's new — Grok 4.6 on Amazon Bedrock - Availability update
GLM-5.3
Z.ai
Direct `glm-5.3` token API listed on the Z.ai pricing table at $1.40/$4.40.
Date note: Same list rates as GLM-5.2. Cached input $0.26 / MTok on the published table.
Z.ai GLM pricing table - Release
GLM-5.3
Z.ai
Date note: Same base as GLM-5.2; gains are post-training. Direct `glm-5.3` token API listed at $1.40/$4.40 from 2026-08-18. Coding Plan access (points quota; 5.2/5.1 requests route to 5.3). Open weights promised about two weeks after launch.
Z.ai GLM-5.3 announcement 
Availability updateDaybreak Red and Daybreak Blue on Amazon Bedrock
OpenAI
Date note: Eligible-customer Bedrock access in US East (N. Virginia). Daybreak Red is GPT-5.6 Cyber; Daybreak Blue is GPT-5.6 Sol with defensive guardrails. Not a new catalog model.
AWS what's new- Limited preview
DeepSeek Harness
DeepSeek
Date note: MIT-licensed v0.1 developer preview (`dsh`). Plugin-based agent runtime with a local web UI via `npx @deepseek-ai/dsh web`. Not a model profile.
DeepSeek Harness product page - General availability
DeepSeek-V4-Pro
DeepSeek
GA checkpoint rolled out on app, web, and API. The API id remains `deepseek-v4-pro`. Peak/off-peak pricing starts 2026-08-16 16:00 UTC.
DeepSeek API changelog 
General availabilityGemini 3.7 Flash
Google
GA Flash workhorse for coding and agents. API id `gemini-3.7-flash`. Introductory paid rates $0.75 / $3.75 per million tokens through 2026-12-31, then $1.50 / $7.50.
Google Gemini 3.7 Flash announcement- Release
Qwen3.8-27B
Qwen
Date note: Apache 2.0 dense 27B VLM, announced with Qwen3.8-Max on 2026-08-03. The Hugging Face repo was created privately on 2026-08-05; the weights were uploaded on 2026-08-13 and the model card finalised on 2026-08-14. Distinct from hosted Qwen3.8-Max.
Qwen3.8-27B on Hugging Face - General availability
Grok 4.6
SpaceXAI
xAI's model card recorded Grok 4.6 as its code and chat default at $2/$6 with a 500K context window; Grok 4.7 replaced it on 2026-09-21.
xAI Grok 4.6 model card 
Availability updateMAI-Thinking-1
Microsoft AI
Date note: Public preview on Microsoft Foundry; closed weights, advertised 256K context and 64,000 output tokens including reasoning. Azure Global Standard meters list $2/$8 per MTok with $0.20 cached input in public regions; government-region meters differ. Preview has no SLA and PTU is unsupported.
Microsoft AI announcement- Release
Qwen3.8-2.4T-A95B
Qwen
Date note: Open-weight checkpoint under the Qwen3.8-Max License. Distinct from the hosted Qwen3.8-Max API.
Qwen3.8-2.4T-A95B on Hugging Face 
Limited previewNVIDIA Nemotron 3.5 Lightning 30B A3B
NVIDIA
NVIDIA's model reference dates the 1.0-preview NVFP4 release.
Date note: OpenMDW 1.1; 30B/3B rounded counts. Vendor BF16 and NVFP4 benchmark columns are distinct from independent board observations.
NVIDIA NIM model reference and version date
Status updateClaude Sonnet 5 $2/$10 made standard
Anthropic
Date note: The previously scheduled 2026-09-01 step to $3/$15 will not occur. Anthropic made $2/$10 the standard price on 2026-08-10.
Anthropic pricing docs
ReleaseMuse Glimmer 30B
Meta
Date note: Open-weight multimodal agentic distill of Muse Spark for ~24–32GB class local hardware; confirm the live HF/license card before redistribution.
Meta Muse Glimmer announcement- Availability update
Ling 3.0 Tiny native-platform launch
InclusionAI
Date note: Native chat/API availability. This date does not establish initial downloadable-weight publication or a later third-party hosting date.
Ant Ling official dated platform changelog 
ReleaseShieldstral
Mistral
Date note: Apache 2.0 safety/moderation model. Covered on open-weights pages; not a general ChatGPT-class catalog row.
Mistral Shieldstral announcement- Release
Qwen3.8-Max
Qwen
Date note: Hosted Qwen3.8-Max API. The open-weight Qwen3.8-2.4T-A95B checkpoint is a separate 2026-08-12 event.
Alibaba Cloud announcement 
Availability updateSakana Namazu v1.0
Sakana AI
Date note: API launch of Sakana's Japanese-focused Kimi K2.6 derivative. Current service excludes EU/EEA, UK and Switzerland; subscription plans do not cover Namazu.
Sakana Namazu API launch
Read the important launches
Gemini 3.7 Flash: the August 2026 coding and agent release
Gemini 3.7 Flash arrived three weeks after 3.6 Flash as Google’s Flash workhorse for coding and agents, with introductory API rates through the end of 2026. Gemini 3.8 Flash (2026-09-02) is newer; compare both on the target workload instead of inferring fit from release order.
Qwen
Qwen3.8-Max: hosted API versus the published open-weight checkpoint
Qwen3.8-Max is live as a hosted API with multimodal extras; the Qwen3.8-2.4T-A95B checkpoint is now on Hugging Face under the Qwen3.8-Max License and is a separate identity.
Meta
Muse Spark 1.2: Meta's August 2026 coding-model update
This August 2026 launch snapshot covers Muse Spark 1.2; Muse Spark 1.3 is the newer model available in the current catalog.
xAI
Grok 4.6: xAI’s August 2026 API release
Grok 4.6 was xAI's August frontier API row, with a 500K context window and published $2/$6 per-million-token rates. Grok 4.7 replaced it on 2026-09-21 at the same price.
DeepSeek
DeepSeek V4-Pro GA: the 0813 checkpoint on the same API id
DeepSeek moved V4-Pro to general availability on 2026-08-13. The API id stayed `deepseek-v4-pro`; hosted billing is peak/off-peak after 2026-08-16.
Meta
Muse Glimmer 30B: Apache-2.0 local agent weights
Muse Glimmer 30B is an open-weight multimodal model described by its release materials as a Muse Spark distill. Hardware fit depends on quantisation, runtime and context; Meta publishes no hosted per-token list for Glimmer.
Meta
Muse Code: Meta's multi-agent terminal agent
Muse Code is Meta's coding agent. Meta describes multi-agent and long-horizon workflows; check the current product page for rollout, platform and account availability.
Z.ai
GLM 5.3: Coding Plan default and separately billed API
GLM 5.3 became the Coding Plan default in August 2026. The hosted token API and open-weight checkpoint have separate access and licensing terms; check Z.ai's current docs for live rates and limits.
Z.ai
GLM 5.3 Flash: Z.ai's lower-cost multimodal model
Z.ai's August 2026 model page describes GLM 5.3 Flash and its hosted API. Treat the earlier 'Ox Alpha' identity attribution as community reporting, not an independently established release fact.