A benchmark of three AI agents on 12 multi-app tasks shows Kimi K3 tied the most expensive model GPT-5.6 Sol at a fraction of the cost, though all three failed cross-app reconcile tasks, highlighting the need for verification in production.
Ran Fable 5, Kimi K3 and GPT-5.6 Sol through 12 multi-step agent tasks over the last couple weeks. Live accounts, not sandboxes. Gmail, Slack, Sheets, Salesforce, HubSpot, GitHub, Linear. No vibe grading. A verifier hits the API after every run and checks what landed in the account against what should have. Setup was Claude Code driving Fable and Kimi, Codex CLI driving GPT-5.6. Same 12 templates, same MCP tool router. Every write tagged so we could clean up after. Scores: Fable 5 - 7/12 Kimi K3 - 7/12 GPT-5.6 Sol - 6/12 Cost per case at list prices, no cache discount, so read these as ceilings: Fable 5 - ~776k tokens, ~$7.76 GPT-5.6 - ~538k, ~$2.69 Kimi K3 - ~463k, ~$1.39 Whole suite came out around $93, $32 and $17. So Kimi tied the most expensive model for a sixth of the money. And the one sitting in the middle on price finished last. Wasn't expecting that going in. The part that bugs me. 5 of the 12 were cross-app reconcile jobs, designed to stress-test the models. Sync ticket state across Gmail/Slack/Sheets. Build a refund ledger out of records scattered across apps. Roster sync, vendor directory. All three failed all five. Zero passes between them. A task only counts if every graded check lands, and the ticket one has 24 of them. One bad merge kills the run. Partial credit told a different story though. GPT-5.6 was usually closest on the ones it lost. 20/24 on ticket where Kimi got 17/24. 12/13 on vendor, 10/13 on refund. Its failures looked like near misses rather than the model going off a cliff. Which is worse for prod, not better.Near misses are easy to miss in production One task split them cleanly. CRM identity dedup, so match contacts across Salesforce and HubSpot, follow canonical-source notes buried in a Gmail thread, hand back a review without touching any records. Fable passed it. Kimi passed it. GPT got 5 of 7 checks and failed. That's the entire gap between 7/12 and 6/12. One task. So where does that leave me? For normal SaaS tool calling the spread is small enough that price decides, and Kimi looks like the better value. For anything that has to land exact state across apps I wouldn't run any of the three unsupervised. Verifier plus a retry loop, and assume it comes back confident and slightly wrong. Caveats before someone yells in the comments. GPT ran on a different harness (Codex vs Claude Code) so account state wasn't identical between runs. It's task-for-task on the same templates, not a clean controlled experiment. Fable and Kimi token counts are runtime tokens normalized per case, GPT's were measured straight. Ratios hold up better than the exact dollar figures. Charts and the per-task breakdown are in the comments if you want them. Anyone got a harness that survives multi-step reconciliation? Wondering if plan-then-verify closes that gap or if it just burns tokens for the same result.
A team shares their experience swapping GPT-4o for Kimi K2.7 in their AI agent, achieving a ~70% reduction in inference costs while noting what functionalities broke or remained intact.
Artificial Analysis benchmarks show OpenAI's GPT-5.6 Sol nearly matches Claude Fable 5 in intelligence at one-third the cost, leads coding agent evaluations, and introduces cache-write pricing.
A developer shares that the best current agent setups use a mix of tools: Kimi, GLM, cloud agents, Fable, and Frontier for orchestration, with DevinAI's ACP running agents locally or in the cloud.
A study introduces DuoBench, a benchmark for evaluating planner-implementer pairs in coding agents. It tests combinations of Kimi K2.7, K2.6, GPT-5.5, and Claude Opus 4.8 on a CPython issue, finding that Kimi K2.7 as an implementer delivers high quality at low cost, outperforming more expensive pairings.