Frontier large-language-model releases from Anthropic, OpenAI, Alibaba (Qwen), xAI, and DeepSeek shipped between mid-July and mid-August 2026
The frontier model wave: five labs, three weeks
Summary
Between mid-July and mid-August 2026, five labs shipped a new flagship model in about three weeks -- Claude Opus 5, GPT-5.6 Sol, Qwen3.8, Grok 4.6, and the general-availability release of DeepSeek V4 Pro. The field converged on one baseline: four of the five disclose a ~1M-token context window, all lean into agentic, long-horizon task execution, and per-token prices held flat or fell even as capability climbed.
By Vindex · August 14, 2026
The round million is now the flagship default, not a premium add-on. GPT-5.6 Sol lists a 1,050,000-token window, Qwen3.8 extends to 1,010,000, and Claude Opus 5 and DeepSeek V4 Pro both sit at exactly 1,000,000. Price moved the other way: Grok 4.6 ships at $2 in / $6 out per million tokens, well under the $5-and-up US flagships.
The ~1M-token flagship window
Disclosed maximum context length of each new mid-2026 flagship (Grok 4.6's page omits a figure)
Claude Opus 5
1,000,000
GPT-5.6 Sol
1,050,000
Qwen3.8-Max
1,010,000
DeepSeek V4 Pro
1,000,000
Source: Anthropic, OpenAI, Qwen, and DeepSeek model documentation (Jul-Aug 2026)
View data as table
| Claude Opus 5 | 1,000,000 | default and maximum; no smaller variant |
|---|---|---|
| GPT-5.6 Sol | 1,050,000 | 922k max input / 128k max output |
| Qwen3.8-Max | 1,010,000 | 262,144 native, extensible to 1.01M |
| DeepSeek V4 Pro | 1,000,000 | 1.6T-param MoE, 49B active |
Labs at a ~1M-token flagship window
4 of 5
Opus 5, GPT-5.6 Sol, Qwen3.8, and DeepSeek V4 Pro all disclose ~1M tokens; Grok 4.6's page omits it
Claude Opus 5 price, in / out per M tokens
$5 / $25
unchanged from Opus 4.8
DeepSeek V4 Pro long-context inference cost
27% of FLOPs
and 10% of KV cache vs DeepSeek-V3.2 at 1M tokens
Grok 4.6 price, in / out per M tokens
$2 / $6
AA Intelligence Index 61, matching GPT-5.6 Sol, at a fraction of the flagship price
The five, model by model
- Claude Opus 5 (Anthropic, July 24): a 1M-token context window as both default and maximum, priced at $5 / $25 per million tokens -- unchanged from Opus 4.8, while Anthropic reports it more than doubles its predecessor on the Frontier-Bench task suite at a lower cost per task.
- GPT-5.6 Sol (OpenAI): the reasoning flagship of the GPT-5.6 family exposes a 1,050,000-token context (922k input / 128k output) at $5 in / $30 out per million tokens.
- Qwen3.8 (Alibaba): an open-weights mixture-of-experts model with 2.4T total / 95B active parameters and context extensible to 1,010,000 tokens; the managed Max variant reports GPQA Diamond 92.6 and SWE-bench Pro 67.7.
- Grok 4.6 (xAI, Aug 12): tuned for long-running agents, it scores 61 on the Artificial Analysis Intelligence Index -- matching GPT-5.6 Sol -- at $2 in / $6 out per million tokens, a fraction of the US flagship price.
- DeepSeek V4 Pro (Aug): a 1.6T-parameter MoE (49B active) with a 1M-token context that, per its technical report, needs only 27% of the inference FLOPs and 10% of the KV cache of DeepSeek-V3.2 at one million tokens; the 0813 build moved from preview to general availability in mid-August.
- The through-line: a ~1M-token window and agentic, long-horizon execution are now the shared baseline, and the newest entrants (Grok 4.6, DeepSeek V4 Pro) compete mainly on price rather than on frontier scores.
Grok 4.6's release page does not state a context-window figure, so the ~1M pattern rests on the four labs that publish one. DeepSeek V4 Pro's Hugging Face model card still describes the series as a "preview"; general availability of the 0813 build is confirmed by DeepSeek's own API changelog, with news reports corroborating an Aug 12-13 GA. All benchmark scores here are vendor-reported and, for these August releases, not yet independently replicated.
Sources(8) ▾
- Anthropic, Introducing Claude Opus 5 (2026-07-24) — Anthropic's own launch announcement for Claude Opus 5. Source for the July 24, 2026 release, the $5 / $25 per-million-token pricing (stated as unchanged from Opus 4.8), and the relative benchmark claims (Frontier-Bench, -AGI 3, OSWorld). Fetched and read in full. anthropic.com
- Anthropic (Claude Platform Docs), What's new in Claude Opus 5 (2026-07-24) — Anthropic's technical model documentation. Source for the 1M-token context window (stated as both default and maximum, no smaller variant), 128k max output tokens, and the $5 / $25 pricing. Fetched and read in full. platform.claude.com
- OpenAI (developer documentation), GPT-5.6 Sol -- API model reference (2026-07-09) — OpenAI's developer API model page for GPT-5.6 Sol. Source for the 1,050,000-token context window (922,000 max input / 128,000 max output tokens) and the $5 input / $30 output per-million-token pricing. Fetched and read in full. developers.openai.com
- Qwen Team, Alibaba (Hugging Face), Qwen3.8-2.4T-A95B (model card) (2026-08-01) — Official Qwen model card for the open-weights base model behind Qwen3.8-Max. Source for the 2.4T total / 95B activated parameter counts, the 262,144-native / 1,010,000-max context, and the reported Qwen3.8-Max benchmarks (GPQA Diamond 92.6, SWE-bench Pro 67.7, Terminal Bench 2.1 86.6). Fetched and read in full. huggingface.co
- xAI, Grok 4.6 (2026-08-12) — xAI's own release page for Grok 4.6. Source for the Aug 12, 2026 release, the long-running-agents focus, the Artificial Analysis Intelligence Index score of 61 (stated to match GPT-5.6 Sol), and the $2 input / $6 output per-million-token pricing. Fetched and read in full. x.ai
- DeepSeek-AI (arXiv:2606.19348), DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence (2026-08-13) — The DeepSeek-V4 technical report. Source for DeepSeek-V4-Pro's 1.6T / 49B-activated parameters, the one-million-token context length, and the efficiency figure (27% of single-token inference FLOPs and 10% of KV cache versus DeepSeek-V3.2 at 1M tokens). Fetched and read. arxiv.org
- DeepSeek-AI (Hugging Face), DeepSeek-V4-Pro (model card) (2026-08-13) — Official DeepSeek-V4-Pro model card, MIT-licensed, linking the technical report (arXiv:2606.19348). The card describes the V4 series as a 'preview'; the GA build (DeepSeek-V4-Pro-0813) is documented in DeepSeek's API changelog. Fetched and read. huggingface.co
- DeepSeek (API documentation), API update: deepseek-v4-pro updated to DeepSeek-V4-Pro-0813 (2026-08-13) — DeepSeek's own API changelog stating that the deepseek-v4-pro endpoint has been updated to the DeepSeek-V4-Pro-0813 build -- the issuer-hosted confirmation of the general-availability release. Fetched and read. api-docs.deepseek.com
Comments
Always open. Logged-in readers can annotate paragraphs in place.
Loading comments…
Between mid-July and mid-August 2026, five labs shipped a new flagship model in about three weeks: Claude Opus 5 (July 24), OpenAI's GPT-5.6 Sol, Alibaba's Qwen3.8, Grok 4.6 (Aug 12), and the general-availability release of DeepSeek V4 Pro. They converged on one baseline: four of the five disclose a ~1M-token context window, all push agentic, long-horizon work, and per-token prices held flat or fell as capability climbed.