Frontier model alternatives for coding: what’s changing
Frontier model alternatives for coding are AI systems that approach the code quality of leading proprietary models on coding benchmarks while cutting inference costs enough to change developers’ cost-performance expectations and purchasing decisions. In this landscape, Zhipu’s GLM-5.2 and Cognition’s SWE-1.7 stand out as two different answers to the same question: how much “frontier” do you need, and what are you willing to pay for it? GLM-5.2 is an open-weight, general-purpose coding workhorse; SWE-1.7 is a proprietary, task-priced specialist focused on code that is safe to merge. In practice, GLM-5.2 suits teams that want control and low per-token bills, while SWE-1.7 suits teams that care most about end-to-end task success on hard software engineering work.
Both models aim to erode single-vendor dependence by offering compelling AI model cost performance outside the classic frontier labs. The choice is no longer “expensive frontier or nothing,” but which cost-performance point along the curve best fits your workloads and risk tolerance.
GLM-5.2: open-weight scale and aggressive inference cost reduction
GLM-5.2 is a 744-billion-parameter mixture-of-experts model that activates about 40 billion parameters per token, cutting active compute while drawing on a much larger knowledge base. That design directly targets inference cost reduction: its API runs at about USD 1.40 (approx. RM6.44) per million input tokens and USD 4.40 (approx. RM20.24) per million output tokens, compared with roughly USD 5 (approx. RM23.00) and USD 25 (approx. RM115.00) for Opus 4.8. For teams stung by surprise token bills, the intelligence-per-dollar ratio is the headline feature: “Companies hit by surprise token bills are now measuring intelligence per dollar, and a cheap model that gets close is a strong answer.”
On coding benchmark comparisons, GLM-5.2 lands within a percentage point of Anthropic’s Opus 4.8 on FrontierSWE and edges past GPT-5.5. It nearly ties Opus on MCP-Atlas and scores 81.0 on Terminal-Bench 2.1 versus Opus at 85.0, a four-point gap. It trails the closed frontier on the hardest from-scratch coding tasks and on Humanity’s Last Exam. The model offers a one-million-token context window aimed at full-project planning, coding, testing, and multi-step loops. Open weights, a permissive MIT license, and the ability to self-host mean buyers gain resilience if access to premium APIs is throttled or revoked. For platform builders, this is the default choice when avoiding lock-in matters more than absolute peak accuracy.
SWE-1.7: task-focused specialist near the frontier
SWE-1.7 takes a different route: a proprietary specialist tuned for hard software engineering tasks rather than a general-purpose open-weight model. Built on top of Moonshot’s Kimi K2.7 base, Cognition’s additional reinforcement learning training delivered large capability gains over an already heavily trained model, pushing back on the idea of a strict post-training ceiling. On Cognition’s FrontierCode 1.1 benchmark, designed to test whether code is actually worth merging, SWE-1.7 scores 42.3%, landing just behind GPT-5.5 at 43.0% and Claude Opus 4.8 at 46.5%, while beating Kimi K2.7 (30.1%), Composer 2.5 (25.6%), and GLM-5.2 (24.5%).
On Terminal-Bench 2.1, SWE-1.7 posts 81.5%, trailing GPT-5.5 (84.2%) and Opus 4.8 (86.9%) but ahead of GLM-5.2 and Composer 2.5. SWE-Bench Multilingual tells a similar story: 77.8% for SWE-1.7 versus 84.4% for Opus 4.8 and 76.8% for GPT-5.5. Cognition prices the model per task: SWE-1.7 costs USD 1.97 (approx. RM9.05) per task on FrontierCode’s Main set and sits on a Pareto-efficient point that pricier models do not occupy. The trade-off is behavior: Cognition notes that SWE-1.7 tends to touch more files and write more tests than strictly required, expanding change scope compared with a more surgical fix. This is ideal for thorough refactors, less so for minimal patches.
Cost-performance, shared caveats, and the move away from single vendors
Taken together, GLM-5.2 and SWE-1.7 show how the coding benchmark comparisons landscape is fragmenting. GLM-5.2 gets very close to frontier leaders on long-horizon tasks like FrontierSWE and MCP-Atlas, with small but consistent gaps on Terminal-Bench 2.1 and difficult from-scratch coding challenges. Independent testing notes that it usually lands just behind Opus on long tasks, and some commentators question whether its scores are tuned to the tests. SWE-1.7 often outruns GLM-5.2 on code quality benchmarks such as FrontierCode, Terminal-Bench 2.1, and SWE-Bench Multilingual, but both models still trail the very top frontier systems on the most demanding workloads.
Where they converge is in reshaping AI model cost performance and inference cost reduction expectations. GLM-5.2’s open weights and much lower token pricing versus Opus 4.8 give buyers a credible frontier model alternative they can download, fine-tune, and self-host. SWE-1.7, by contrast, shows that an application company can build a proprietary model on rented and mixed infrastructure, compress weight updates by more than 99%, and still price it aggressively enough to influence the wider market. As one source notes, for buyers who cannot afford to have a tool switched off, a model no agency can revoke starts to look safer, while SWE-1.7’s aggressive pricing shows that even proprietary stacks must now compete on value. Both forces push developers away from relying on a single vendor and toward diversified model portfolios.
| Aspect | GLM-5.2 | SWE-1.7 |
|---|---|---|
| Access model | Open-weight, permissive MIT license; downloadable, fine-tunable, and self-hostable | Proprietary model operated by Cognition for its products; no open weights |
| Pricing style | Approx. USD 1.40 (approx. RM6.44) /M input tokens and USD 4.40 (approx. RM20.24) /M output tokens; ~1/5 Opus 4.8’s listed rates | Approx. USD 1.97 (approx. RM9.05) per task on FrontierCode Main; positioned on a favorable score-vs-cost Pareto curve |
| Headline coding benchmarks | Near Opus 4.8 on FrontierSWE; nearly ties on MCP-Atlas; 81.0 on Terminal-Bench 2.1 vs Opus at 85.0; 99.2% on AIME 2026 math | 42.3% on FrontierCode 1.1 vs 46.5% for Opus 4.8 and 43.0% for GPT-5.5; 81.5% on Terminal-Bench 2.1; 77.8% on SWE-Bench Multilingual |
| Strengths | Low token cost, huge context window for large codebases, self-hosting, strong long-horizon planning and coding | Near-frontier task performance, high merge-worthy code quality, competitive score-per-dollar, strong results on multilingual and end-to-end coding tasks |
| Noted caveats | Trails closed frontier models on hardest from-scratch coding tasks and Humanity’s Last Exam; concerns about benchmark tuning | Tends to touch more files and write more tests than necessary, expanding change scope beyond minimal fixes |
Buy if / Skip if
- Buy the GLM-5.2 if you want open weights, self-hosting, and strong coding benchmark comparisons without paying frontier token prices.
- Skip the GLM-5.2 if your workloads demand absolute best-in-class performance on the hardest from-scratch coding and exam-style tasks.
- Buy the SWE-1.7 if your priority is near-frontier, merge-ready code on complex software tasks at a task-based price that undercuts top-tier APIs.
- Skip the SWE-1.7 if you need strict, minimal edits to existing codebases where touching extra files or adding many tests is a problem.
- Buy the GLM-5.2 if reducing single-vendor dependency and keeping a model that cannot be switched off by a third party is essential to your risk planning.
- Skip the GLM-5.2 if you prefer a fully managed, proprietary service and do not wish to deal with model hosting or fine-tuning pipelines.
- Buy the SWE-1.7 if you are building long-horizon coding agents and care most about end-to-end task success rather than raw per-token flexibility.
- Skip the SWE-1.7 if open-source licensing, on-prem deployment, or deep model customization are non-negotiable requirements for your stack.
