Learn AI · News
Model Release Changelog — August 2026
Monthly roundup of frontier model releases, pricing changes, and open-weight drops — GPT-5.2, Sonnet 4.7 500K context, DeepSeek V3.2 GA, Llama 4.2 multimodal, and what to do about each.
Edited by The AIKnowHub team · Editorial teamKey takeaways
- 1GPT-5.2 (Aug 5): 12% reasoning quality bump, 30% fewer tool-call format errors, same pricing as GPT-5.1.
- 2Sonnet 4.7 500K preview (Aug 11): Anthropic extends coding-agent context without Opus pricing — request access via API waitlist.
- 3DeepSeek V3.2 GA (Aug 18): open weights, matches GPT-5.1 on MATH, 15% faster inference vs V3.1 preview.
- 4Llama 4.2 multimodal (Aug 24): image-in/text-out, strong on diagrams and UI screenshots — runs on single H100.
- 5Trend: August was an infrastructure month — bigger context, open reasoning GA, multimodal open weights. Few headline price cuts.
Recurring format: Monthly roundup of frontier closed-model releases, open-weight drops, pricing changes, and practical recommendations. June · July · August 2026 (this page)
August 2026 at a glance
| Date | Release | Category | Impact |
|---|---|---|---|
| Aug 5 | GPT-5.2 | Closed | High — agent reliability |
| Aug 11 | Sonnet 4.7 500K preview | Closed | High — large-repo coding |
| Aug 18 | DeepSeek V3.2 GA | Open | High — open reasoning |
| Aug 24 | Llama 4.2 multimodal | Open | Medium — vision tasks |
| Aug 27 | Mistral Pixtral Large 2 | Open | Medium — document OCR |
| Aug 30 | Gemini 2.5 Flash price cut | Closed | Low — throughput workloads |
Frontier closed models
OpenAI — GPT-5.2 (August 5)
Incremental but agent-relevant.
What shipped:
- GPT-5.2 — reasoning quality bump, improved tool-call JSON reliability.
- Realtime API support for GPT-5.2 (voice agent backends).
- Deprecated GPT-4.1 from ChatGPT Team — migration notices sent.
Benchmark highlights:
- MATH: 95.8% (up from 94.2% on GPT-5.1).
- Tool-call valid JSON rate: 97.2% (up from 92.1% on GPT-5.1).
- Agent task completion (τ-bench subset): +12% over GPT-5.1.
Pricing: Same as GPT-5.1 — $2.50/$10 per 1M tokens. No change to Mini or Pro tiers.
What to do: Upgrade agent backends first — especially if you log tool-parse failures. Keep Mini for classification. See Best AI Model for Coding.
Anthropic — Sonnet 4.7 500K context preview (August 11)
Bigger context without Opus pricing.
What shipped:
- Sonnet 4.7 — 500K token context window (preview, waitlist).
- Claude Code beta flag
CLAUDE_CODE_MAX_CONTEXT=500000for eligible accounts. - Prompt caching extended to 500K sessions — 90% discount on cached prefix.
Benchmark highlights:
- Long-codebase navigation (internal eval): +18% correct file identification vs 200K default.
- Latency at 500K: 2.3x slower first token — use deliberately, not by default.
- Quality at 400K+ tokens: degrades on middle sections — same "lost in the middle" pattern as Gemini.
Pricing: Preview pricing matches standard Sonnet $3/$15. No GA date announced.
What to do: Request preview access if you run multi-hour agent sessions on monorepos. Keep RAG for doc Q&A — 500K is for active editing sessions, not retrieval replacement. See chunking strategies for RAG.
Google — Gemini 2.5 Flash price cut (August 30)
Minor end-of-month move.
What shipped:
- Gemini 2.5 Flash — 20% price cut on input tokens.
- Batch inference discount increased to 60% off for Flash workloads.
Pricing: Flash input $0.05 → $0.04 per 1M tokens. Pro and Ultra unchanged.
What to do: Shift high-volume extraction and classification to Flash if you're on Pro for cost reasons only. See Cheapest AI APIs.
Open-weight releases
DeepSeek — V3.2 GA (August 18)
July's preview delivered.
What shipped:
- DeepSeek V3.2 — full open weights (MoE architecture, 671B total / 37B active).
- MIT license on model weights — commercial use allowed.
- vLLM and SGLang day-one support; Ollama build within 72 hours.
Benchmark highlights:
- MATH: 94.6% — matches GPT-5.1 (94.2%).
- GPQA: 81.3% — behind Opus 4.7 (84.2%) but ahead of all prior open models.
- Inference: 15% faster than V3.2 preview on H100 clusters.
What to do: Best open reasoning model available. Self-host for internal research, code review, and analysis workloads. Pair with guardrails for any user-facing surface. See local models vs API.
Meta — Llama 4.2 multimodal (August 24)
Image-in, text-out — open weights.
What shipped:
- Llama 4.2 Vision 11B and 90B — diagram, chart, screenshot, and UI understanding.
- Llama 4.2 90B text-only — minor bump over Llama 4.1 70B on coding.
- License unchanged — Meta custom license with commercial thresholds.
Benchmark highlights:
- Chart/diagram QA (AI2D): 84.7% on 90B Vision — competitive with Gemini Pro on static images.
- Screenshot-to-code (internal Meta eval): 71% functional HTML — useful for prototyping, not production codegen.
- Runs on 1x H100 (90B quantized) or M3 Max 64GB (11B).
What to do: Use 11B for local diagram explanation and support screenshot triage. Don't replace coding agents — pair with Llama 4.1 70B or API models for multi-file edits.
Mistral — Pixtral Large 2 (August 27)
Document-focused vision model.
What shipped:
- Pixtral Large 2 — 124B multimodal, optimized for OCR and document layout.
- Mistral OCR API — hosted service for PDF → structured markdown.
Benchmark highlights:
- PDF table extraction F1: 91.2% — best open OCR pipeline this month.
- Handwriting recognition: moderate — don't use for medical forms without human review.
What to do: Strong for document ingestion pipelines feeding RAG. Compare against Gemini Pro on 20 real PDFs before switching.
Pricing changes
| Model | Old price (per 1M tokens) | New price | Change |
|---|---|---|---|
| Gemini 2.5 Flash input | $0.05 | $0.04 | -20% |
| GPT-5.2 input | $2.50 | $2.50 | — (new model, same tier) |
| GPT-5.2 output | $10.00 | $10.00 | — |
| Claude Sonnet 4.7 input | $3.00 | $3.00 | — |
| DeepSeek V3.2 (self-hosted) | — | GPU cost only | open weights |
Trend: Quiet month for pricing. July's Pro-tier cuts held. Flash tier still inching down. Value shifted to capability — context, multimodal, open reasoning — not cheaper tokens.
Update dashboards — see LLM Cost Optimization.
What to do this month
If you build coding agents
→ Request Sonnet 500K preview if monorepo context limits your sessions. Default stays Sonnet 4.7 200K for latency. Upgrade OpenAI agents to GPT-5.2 if tool-parse errors appear in logs.
If you self-host
→ Download DeepSeek V3.2 for reasoning workloads. Add Llama 4.2 Vision 11B for diagram/screenshot tasks. Keep Llama 4.1 70B for pure coding until you benchmark 4.2 90B text.
If you run document RAG
→ Benchmark Pixtral Large 2 or Llama 4.2 Vision on your PDF corpus before paying for vision API calls. May cut ingestion cost significantly.
If you optimize spend
→ Move high-volume Flash-eligible work to Gemini 2.5 Flash after the 20% cut. No other mandatory routing changes this month.
If you're catching up on the series
→ June (Sonnet 4.7, GPT-5.1) → July (Opus GA, Qwen 3) → August (this page).
September and October changelogs
Sonnet 500K GA, GPT-5.2 Mini, Grok 3.5 API — see September 2026. GPT-5.3 preview, Opus 4.8 preview, Qwen 3.5 API — see October 2026.
Recurring series. June 2026 · July 2026 · August 2026 (this page) · September · October · RSS
Common misconceptions
The wrong-but-common takes worth correcting.
Myth
GPT-5.2 obsoletes GPT-5.1 overnight.
Reality
GPT-5.2 improves reasoning and tool formatting — meaningful for agent loops, marginal for simple chat. Migrate agent workloads first; leave classification on Mini.
Myth
Sonnet 500K means you can drop RAG for codebases.
Reality
500K helps for large-repo agent sessions, but retrieval quality still degrades in the middle of huge contexts. Hybrid approach — RAG for lookup, long context for active edit session — still wins.
Myth
DeepSeek V3.2 is only for China-region teams.
Reality
Open weights under permissive license — any team with GPU capacity can self-host. Western API providers still lead on safety tuning and enterprise SLAs for customer-facing chat.
Real-world use cases
Agent loop reliability
GPT-5.2's tool-format improvements reduce parse-error retries in production agent pipelines.
Large-repo coding sessions
Sonnet 500K preview lets you test whole-monorepo context without stuffing retrieval chunks.
Self-hosted reasoning
DeepSeek V3.2 GA is the month's pick for teams exiting API spend on internal reasoning tasks.
Diagram and UI understanding
Llama 4.2 multimodal for architecture diagrams, wireframes, and screenshot-based support tools.
Frequently asked questions
Watch
Hand-picked videos from official + trusted channels. Opens in a new tab.
Related on AIKnowHub
Concept
Model Release Changelog — July 2026
Monthly roundup of frontier model releases, pricing changes, and open-weight drops — Claude Opus 4.7 GA, GPT-5 Pro price cut, Gemini 2.5 Ultra, Qwen 3 family, and what to do about each.
Concept
Model Release Changelog — June 2026
Monthly roundup of frontier model releases, pricing changes, and open-weight drops — Claude, GPT, Gemini, Llama, Mistral, and DeepSeek. First edition of a recurring series.
Concept
Guardrails & Output Validation
LLMs don't come with guarantees. Guardrails and output validation are how you enforce safety, structure, and policy on probabilistic outputs before they reach users.
Comparison
Best AI Model for Coding
Which model actually writes the best code? Benchmarked against real tasks — refactors, bug fixes, new features, code review.
Comparison
GPT vs Claude vs Gemini
An honest head-to-head of the three frontier models — where each one wins, where each one loses, and what to pick.