Versus Anthropic's Claude Family
Pricing and production maturity. Claude Opus 5 has been running production agentic-coding workloads since July 2026 at $5 per million input tokens — half of Astra's $10 — while scoring 96.0% on SWE-bench. By the time Astra launched in September 2026, many of its benchmark numbers (and Gemini 3.8 Flash's) were only days old, whereas Opus 5 had already racked up two months of real-world production mileage, meaning some of its benchmark scores have been validated over a longer stretch of actual use.
Astra's $10/$50 pricing lands exactly on another Anthropic model, Claude Fable 5.1 — a sign that the leading labs may be converging on a similar price band for their flagship offerings.
Mixed benchmark results. Specifically:
- On Humanity's Last Exam with tools, Astra scores 57.2%, trailing Claude Fable 5.1's 65.0% and Claude Opus 5's 63.6%.
- On FrontierCode 1.1 Main, Astra's 53.3% is essentially tied with Claude Fable 5's 53.5% and Opus 5's 53.4%.
- On DeepSWE v1.1, Astra's 74.1% leads Gemini 3.8 Flash's 73.8% and Fable 5's 69.9%.
- On Terminal-Bench 4.0, Astra's 57.7% edges out Claude Fable 5.1's 55.8%.
- On the ExploitBench cybersecurity benchmark, Astra's 100% clearly outpaces Claude Opus 5's 70%.
- On BrowseComp (a web-research benchmark), the three are nearly tied: Astra 91.5%, Opus 5 90.8%, Sol 90.4%.
On Artificial Analysis's composite scoring, Fable 5.1 leads Astra 70 to 67 on the Coding Agent Index, and 66 to 61 on the Intelligence Index (where Astra ties Sol). Under this particular independent evaluation framework, Claude Fable 5.1 currently holds a modest edge over GPT-6 Astra on both overall intelligence and coding-agent performance.
Different philosophies on context windows. The three companies have taken markedly different approaches to context window design. Gemini 3.7 supports over 2 million tokens — the "volume king" — well suited to ingesting massive, uncurated raw material, including hours of video. GPT-6 Astra's roughly 1.05-million-token window favors dense state retention and AST-level information persistence. Claude Opus, by contrast, deliberately caps its window between 200,000 and 500,000 tokens to maximize "attention density" — when cross-referencing complex variables across 50 separate modules, its multi-hop reasoning accuracy stays essentially flat from token 100,000 to 400,000, meaning it experiences the least "context rot" of the three.
The practical takeaway: Gemini is the better fit for ingesting huge volumes of unstructured data in one pass; Claude's approach is more reliable for precise multi-step reasoning within a smaller but denser window; and Astra sits in between, trying to balance capacity against density.
Computer-use comparison. On OSWorld 2.0, Claude Opus 5 scores 70.2%, offering reliable visual navigation but a noticeably slower interaction loop — Anthropic tends to insert safety checkpoints before key actions, frequently asking for user confirmation. Astra, by contrast, scores 72.6% on the same benchmark with a shorter average time per task, reflecting OpenAI's greater emphasis in this generation on fluid, autonomous execution.
Versus Google Gemini
Google's Gemini 3.8 Flash is the cheapest of the three closed-source flagships, priced at just $0.75 per million input tokens — far below Astra's $10 and Claude Opus 5's $5. It also boasts the largest context capability of the three: according to Google DeepMind's official model card, it supports a 1-million-token context window, a 64,000-token output limit, and a March 2026 knowledge cutoff.
That low price and large capacity come with a tradeoff, though: on specialized coding and terminal-task benchmarks like Terminal-Bench 4.0, Gemini 3.8 Flash scores notably lower (around 19%), well behind Astra's 57.7% and Claude Fable 5.1's 55.8%. This positions Gemini 3.8 Flash more as a value-and-scale play for large-volume data processing than a top contender for elite agentic-coding or terminal-operation capability — which still makes it highly attractive for budget-conscious teams focused on information retrieval and summarization.
Versus China's Open-Weight Camp
From mid-2026 through September, China's open-weight model ecosystem experienced a concentrated wave of releases, putting substantial pricing and performance pressure on closed-source flagships, including GPT-6 Astra.
- Kimi K3 (Moonshot AI). The largest open-weight model released to date, at 2.8 trillion parameters, initially available API-only with weights still pending release. On the Artificial Analysis Intelligence Index, K3 ranked fourth globally — the highest position ever recorded by an open-weight model, surpassing the then-current Claude Opus 4.8. On the specific SWE-bench Verified coding benchmark, though, K3 scored lower than DeepSeek V4 Pro (80.6%) and GLM's previous generation (77.8%), suggesting its coding strength shows up more clearly on other measures than on that particular test. Moonshot AI committed to releasing the weights under a Modified MIT license by the end of July 2026, with an attribution clause that only triggers above 100 million monthly active users — effectively irrelevant for most commercial applications.
- DeepSeek V4 Pro. The veteran of the price war. Its SWE-bench Verified score is roughly on par with the then-current Claude Opus 4.7, and it remains one of the best-documented, most production-ready open-weight endpoints available. DeepSeek has recently raised its peak-hour pricing as it has iterated, though it still runs significantly below closed-source flagship pricing.
- GLM-5.3 (Zhipu AI). The newest arrival in this wave. GLM-5.3 topped the CyberGym cybersecurity benchmark and reportedly beat an upgraded DeepSeek V4-Pro on Terminal-Bench with a score of 28.3, earning recognition in some evaluations as the new open-coding benchmark leader. GLM's monthly subscription pricing starts as low as about $12.60, one of the most cost-effective options among the three.
- MiniMax and others. In earlier comparative testing, models like MiniMax M2.7, the Qwen family, and Step also competed on the same field, generally landing in the mid-tier — a reminder that China's model ecosystem isn't a simple three-way race between DeepSeek, Kimi, and GLM, but a multi-tiered landscape.
A steep pricing gap. Based on vendor-direct API rates as of early September 2026, GPT-6 Astra's output-token cost is 8 times Alibaba's Qwen 3.8-Max, 40 times DeepSeek V4.1 Flash, and 100 times GLM 5.3 Flash. That gap means open-weight models already offer an extremely compelling cost-performance ratio for budget-conscious teams whose workloads overlap heavily with these models' strengths — routine coding tasks and terminal operations, for example.
Differences in ecosystem fit. Different open models fit differently into developer ecosystems: developers using the Claude Code harness tend to prefer GLM, Codex users tend to prefer DeepSeek, and Kimi K3 has the strongest multimodal understanding of the three. The right open-weight choice depends not just on benchmark scores, but on compatibility with your existing development toolchain.
No Single Model Wins Across the Board
Pulling this together, a few conclusions stand out:
- The closed-source flagships trade wins with each other. OpenAI claims near-perfect scores on abstract reasoning benchmarks like ARC-AGI-3 (99.9%) and FrontierMath (98%); Anthropic's Claude Fable 5.1 leads on Artificial Analysis's composite Intelligence Index and Coding Agent Index; Google's Gemini carves out its own niche with a massive context window and highly competitive pricing.
- China's open-weight models are now genuinely competitive. On specialized benchmarks — cybersecurity, terminal tasks, select coding evaluations — models like DeepSeek, Kimi, and GLM can go head-to-head with closed-source flagships, at a fraction (sometimes a tiny fraction) of the per-token cost.
- The right choice comes down to task and budget. For work requiring high trust and strict security review — sensitive cybersecurity testing or enterprise-grade agentic automation — closed-source flagships (especially Astra and Claude) still hold an advantage. For routine coding and content-generation work where cost sensitivity is high, open-weight models offer a cost-performance ratio that's nearly impossible to match.
Data in this article is compiled from public sources as of mid-September 2026. Benchmark scores, pricing, and rankings for all models mentioned may change with future updates — consult each vendor's official announcements for the latest information.
For a narrative look at what GPT-6 Astra can actually do once you move past the benchmark table, see how it operates inside Blender as an agentic 3D tool, or browse all blogs for the rest of our coverage.