It's coming up on a year since the models actually changed. Not incremental, a real shift in what AI can do for engineers. In that year the whole landscape has rewritten itself.
The numbers tell it better than words. August 2025: Claude 3.5 Sonnet at 200K context and 49% SWE-bench. August 2026: Claude Fable 5 at 1M context and 95% SWE-bench. Twelve months and the headline numbers more than doubled.
The Model Landscape (August 2025)
Three models mattered for coding.
Claude 3.5 Sonnet (Anthropic) — 200K context, $3/$15 per MTok. 92% HumanEval, 49% SWE-bench Verified. This was the SOTA coding model. No agent mode, no sub-agents, no computer use. The 200K context was considered generous.
GPT-4o (OpenAI) — 128K context, $2.50/$10 per MTok. 90.2% HumanEval, 38.1% SWE-bench. Chat interface only. No dedicated coding agent.
Gemini 1.5 Pro (Google) — 1M context (huge for its time), $3.50/$10.50 per MTok. 84.1% HumanEval. The 1M context was its standout feature, but coding scores lagged.
DeepSeek V3 existed but wasn't competitive on coding benchmarks. No agent-specific benchmarks existed yet. SWE-bench Verified was the hardest test.
Pricing: $2.50-$3.50 per MTok input for frontier models. Agentic coding wasn't a category.
The Model Landscape (August 2026)
Five model families compete, and the numbers have separated into tiers.
Claude Fable 5 (Anthropic, June 2026) — 1M context, $10/$50 per MTok. 95% SWE-bench Verified, 80.3% SWE-bench Pro. The undisputed coding leader. Also leads on computer use at 83.4% on OSWorld-Verified. Access is restricted.
Claude Opus 4.8 (Anthropic) — 1M context, $5/$25 per MTok. 88.6% SWE-bench Verified, 69.2% SWE-bench Pro, 78.9% Terminal-Bench 2.1. The production workhorse. Nearly doubled the SWE-bench score from Sonnet 3.5 in 12 months.
GPT-5.5 (OpenAI, via Codex CLI) — 1M context, $2/$10 per MTok. 88.7% SWE-bench Verified, 83.4% Terminal-Bench 2.1. The Terminal-Bench leader. Best value in the frontier tier.
DeepSeek V4 Flash 0731 — 1M context, $0.14/$0.28 per MTok. 82.7% Terminal-Bench 2.1 (official, vendor-reported). MIT licensed. The price disruptor. At $0.14/M it costs 1/70th of Fable 5 for input.
DeepSeek V4 Pro — 1M context, $0.44/$0.87 per MTok. 80.6% SWE-bench Verified. Open weights (MIT-ish license).
Gemini 3.5 Flash (Google) — 1M context, $0.15/$0.60 per MTok. 83.6% MCP Atlas. Fastest time-to-first-token in the frontier tier.
Kimi K3 (Moonshot AI, July 2026) — 1M context, $3/$15 per MTok. 88.3% Terminal-Bench 2.1, 93.5% GPQA Diamond, #1 on BrowseComp and Program Bench. Open weights (2.8T MoE, released July 26). The strongest open-weight coding model, competitive with frontier closed models at half the price.
Kimi K2.7 Code (Moonshot AI, June 2026) — 256K context. Agent Swarm capability up to 300 sub-agents. Coding specialist tuned for the Kimi Code CLI, using 30% fewer thinking tokens than K2.6.
What the numbers say
SWE-bench Verified went from 49% (Sonnet 3.5) to 95% (Fable 5). 46 points in 12 months. The same benchmark. Context windows went from 200K standard to 1M standard. Every frontier model ships 1M now. Pricing collapsed at the low end: $3/MTok to $0.14/MTok for capable coding. And entirely new benchmarks emerged: Terminal-Bench (agentic coding), MCP Atlas (tool use), OSWorld (computer use). All measuring capabilities that didn't exist as categories a year ago.
What hasn't changed
Despite the model gains, code review is still the bottleneck. The PR queue gets longer. Architecture decisions still need a person. The models are better at implementing within patterns but not at deciding which patterns to use.
Takeaway
The models doubled their coding benchmarks, quintupled their context windows, dropped prices by 95% at the low end, and created an entirely new agentic capability category that didn't exist 12 months ago. The industry trajectory is clear: implementation is being commoditised. Architecture and judgment are the remaining moat.
Top comments (0)