The model stopped being the whole product. In 2026 the thing that decides whether your agent finishes the job is the harness wrapped around it: the tools, the context discipline, the verification, the permissions.
The one idea
agentic performance = model × harness × tools × context discipline × verification × permissions
Every term is a multiplier, so a zero anywhere zeroes the product. A frontier model with no verification loop produces confident, plausible, unfinished work. A perfect harness driving a model that cannot re‑plan after a surprise stalls on the first failed tool call.
This is why a mid‑tier model inside a strong harness routinely beats a flagship model pasted into a thin chat box, and why the interesting engineering moved from prompt writing to harness building.
Prices are approximate API dollars per million tokens, input and output. Treat a 1M‑token window as a capacity specification, not a promise that the model still respects your constraint at token 700,000.
| Model | Context / output | Cost in / out | Agentic read | Reach for it when |
|---|---|---|---|---|
| Anthropic | ||||
| Claude Fable 5 | 1M / 128K | $10 / $50 | Maximum horizon. Strongest at planning, recovery, and holding one coherent thread across days. Ruinous as a routine worker. | Multi‑day autonomous work, gnarly migrations, reviewing another agent’s output. |
| Claude Opus 5 | 1M / 128K | $5 / $25 | Near‑Fable judgement at half the price. Good in messy repos and underspecified requirements. | Hard engineering, architecture calls, incident analysis, final review. |
| Claude Sonnet 5 | 1M / 128K | $2 / $10$3 / $15 from Sept 1 | The best everyday Claude. Strong code editing, tool selection, and instruction adherence, especially inside Claude Code. | Default interactive coding and most knowledge‑work agents. Escalate the few loops it cannot close. |
| Claude Haiku 4.5 | 200K / 64K | $1 / $5 | Fast and cheap. Weak at re‑planning once something surprises it. | Extraction, classification, test generation, narrow subagents whose output is checked by code. |
| OpenAI | ||||
| GPT‑5.6 Sol | 1.05M / 128K | $5 / $302× in, 1.5× out above 272K | Strongest general agent model. Excellent across coding, computer use, research, and heterogeneous tool work. | Hard general‑purpose agents and consequential coding, especially in Codex. |
| GPT‑5.6 Terra | 1.05M / 128K | $2.50 / $152× in, 1.5× out above 272K | The cost and performance sweet spot. Slightly less margin than Sol on ambiguous long‑horizon work. | Most production agent calls and ordinary Codex sessions. |
| GPT‑5.6 Luna | 1.05M / 128K | $1 / $6 | Cheap and capable inside tight constraints. Needs schemas, short horizons, deterministic checks. | Extraction, classification, routing, search workers. |
| Gemini 3.6 Flash | 1M / 64K | $1.50 / $7.50 | Startling value. Native multimodality, search, file and URL handling, code execution. Less trustworthy than top deliberative models when the task is underspecified. | Document and web‑heavy agents, large corpora, high‑throughput tool work. |
| Gemini 3.1 Pro | 1M / 65K | $2 / $12 under 200K | Better deliberate reasoning and custom‑tool handling than Flash. Still a preview surface. | Google‑centric agents that need more planning than Flash gives. |
| Gemini 3.5 Flash‑Lite | 1M / 65K | $0.30 / $2.50 | Very cheap, very fast, no business owning a complex job on its own. | Bulk parsing, routing, tagging, disposable verified workers. |
| Family | Context | Licence | Assessment | Reach for it when |
|---|---|---|---|---|
| Qwen3.6 | 262K typical | Apache‑2.0 | Best practical open default. The 27B dense and 35B‑A3B checkpoints are genuinely deployable and good at coding with interleaved tool use. | Local coding, private enterprise agents, high‑throughput subagents. |
| DeepSeek V4 | 1M / 384K | Hosted API | Extraordinary economics. Flash near $0.14 / $0.28, Pro near $0.435 / $0.87. Weigh provider trust and data residency separately from quality. | Cost‑dominated pipelines and verified background workers. |
| Kimi K2.7‑Code | 256K | Modified MIT | 1T total parameters, 32B active. Strong inside code‑agent loops, expensive to self‑host at full scale. Trims thinking tokens versus K2.6, which directly lowers long‑run cost. | You want a powerful self‑hosted coding specialist and own the hardware. |
| GLM‑5.2 | 1M | MIT | One of the strongest open picks for long‑horizon agents. The 753B scale is operationally demanding and serving quality dominates. | Large organisations with serious inference infrastructure. |
| Mistral Small 4 / Devstral 2 | Varies | Apache‑2.0Devstral 2 123B is modified MIT | Attractive for European and on‑prem deployment. Devstral Small 2 at 24B is practical. Still behind the closed frontier at recovering from complex tool failures. | Sovereign deployments, local IDE agents, moderate hardware. |
| Llama 4 Scout / Maverick | Nominal 10M (Scout) | Custom Meta | Broad ecosystem, no longer the performance default for autonomous coding. Do not assume reliable recall across that advertised 10M span. | Existing Llama estates, sovereignty requirements, fine‑tuning ecosystems. |
The harnesses
This is where the last twelve months of progress went. Skills, hooks, MCP, subagents, sandboxes, resumable sessions and native verification now create bigger practical differences than the benchmark gaps between the models they wrap.
A good harness lifts a cheaper model by supplying precise repository maps and just‑in‑time context, constraining tools and validating arguments, keeping a compact decision ledger, running tests automatically and reading the failures, checkpointing before anything risky, and escalating only after repeated failure.
This is the whole reason Sonnet in Claude Code beats a flagship model in a chat window. The mechanics are handled, so the model only has to think.
The expensive model wins when the bottleneck is judgement rather than mechanics: ambiguous requirements, architecture, novel debugging, recovery after several consecutive failures, cross‑domain synthesis, adversarial review, or staying coherent for hours.
No amount of scaffolding substitutes for a model that can notice it was wrong and change plan.
One caveat that costs people real money: do not put a weak model in the orchestrator seat just because the harness is good. Routing and recovery are precisely where cheap models make small mistakes that compound.
One agent, one explicit objective, a small tool set. A single coherent context beats multi‑agent communication overhead for most work. Add subagents only when the work is genuinely parallelizable: repository exploration, retrieval, test execution, security review, competing proposals. Three to eight workers is usually plenty. The orchestrator holds the acceptance criteria and merges. Workers return artifacts, evidence and conclusions, never whole transcripts.
If the stages are stable, ingest, extract, validate, enrich, approve, publish, then build a pipeline and put models inside individual stages. Do not ask an autonomous agent to rediscover your business process on every run. LangGraph when branching state and replay are central. Temporal when side effects, waiting, retries, compensation or month‑long execution matter. Running both is reasonable.
Keep a compact working record: goal and acceptance tests, non‑negotiable constraints, current plan and completed steps, decisions with their evidence, failed approaches, open risks. Retrieve big source material just in time. Compaction should preserve decisions and provenance, not conversational tone.
Store durable facts with source, timestamp, confidence, owner and expiry. A large vector store full of old agent utterances is not memory, it is uncurated contamination.
Tests, linters, type checkers, schema validation, policy checks and simulators wherever they exist. For high‑risk work add an independent reviewer that sees the original requirements and the final artifact, not the first agent’s reasoning trace. Use deterministic metrics for exact checks like tool correctness, and reserve model judges for things that genuinely need judgement. Calibrate any judge against traces you have labelled yourself.
Constrain filesystem and process access. Deny network and secrets unless asked. Short‑lived credentials, explicit allowlists. Require a human for production deploys, payments, customer communication, permission changes, data deletion and irreversible migrations. Human‑in‑the‑loop has shifted from approving every line to approving the plan.
Measure task success, pass@1, recovery rate, invalid tool calls, human interventions, latency, tokens, dollars and side effects, against a pinned combination of model, prompt, harness, tools and sandbox. A benchmark score for a raw model tells you very little about your system. Feed failing production traces back in as new eval cases so the suite grows from reality.
Economics
Agentic loops will happily bankrupt you. Caching, tiering and hard caps do most of the work. Be sceptical of any blog that quotes you a percentage saving: almost none of them define a baseline, a workload or a quality target, and caching and routing overlap so their savings cannot simply be added together.
Discount on cached prompt prefixes, and the one number here worth designing around. Order messages so the stable part (system prompt, tool schemas, repo summary) never moves.
The GPT‑5.6 cliff. Cross it and the entire request re‑prices at 2× input and 1.5× output, not just the overflow.
What OpenAI reports Codex actually costs per developer per month. For interactive coding this, not the per‑token rate, is the number that decides things.
Reported cost of one automated code review. Reviewing agent output can cost more than generating it, and deserves its own budget line.
API price tables are the wrong lens for interactive coding. Claude Pro is $20 a month, Max is $100 or $200, and included use is bounded by five‑hour and weekly limits rather than per‑token rates. Compare monthly allowance, throttling behaviour and ergonomics first. Token pricing only becomes the primary lens once you are running agents against the API at volume.
Tokens, dollars, wall time, tool calls, retries, fan‑out. Stop after repeated identical failures and escalate, rather than letting a stochastic loop keep paying for the same mistake.
Subagents earn their keep when they parallelize independent expensive searches, or keep noisy context out of the main thread. They lose money when tasks are sequential, when workers need the same full context anyway, or when several agents edit the same files. Every subagent duplicates setup tokens and creates a merge obligation.
| Task | Model | Harness |
|---|---|---|
| Everyday coding in an existing repo | Claude Sonnet 5 | Claude Code |
| Hard cross‑repository migration | GPT‑5.6 Sol or Opus 5 | Codex or Claude Code |
| Multi‑day investigation or autonomous repair | Claude Fable 5 | Claude Code or Agent SDK, with checkpoints and gates |
| IDE‑first pair programming | Sonnet 5 or GPT‑5.6 Terra | Cursor, or Windsurf |
| Transparent Git‑native workflow | Sonnet 5, Qwen3.6, Devstral | Aider |
| Open or self‑hosted coding platform | Qwen3.6‑35B‑A3B or Kimi K2.7‑Code | OpenHands |
| High‑volume document and multimodal work | Gemini 3.6 Flash | Google API runtime |
| Cheap verified background workers | DeepSeek V4 Flash, Flash‑Lite, Haiku 4.5, Luna | Pipeline, LangGraph, or Agents SDK |
| OpenAI‑native customer or service agent | GPT‑5.6 Terra | OpenAI Agents SDK |
| Claude‑native workspace or research agent | Claude Sonnet 5 | Claude Agent SDK |
| Stateful branching process | Terra, Sonnet, or Gemini Flash | LangGraph |
| Long‑running consequential workflow | Terra or Sonnet | Agents SDK or LangGraph above Temporal |
| Turnkey asynchronous tickets | Service selected | Devin |
| Customizable bring‑your‑own‑model IDE agent | Qwen3.6, Sonnet, or Terra | Cline or Roo |
| Fast role‑based automation prototype | Terra or Sonnet | CrewAI |
Skills, hooks, MCP, subagents, background execution, sandboxes, resumable sessions and repository instructions now separate stacks more than a few benchmark points separate the models inside them.
Retrieval friction dropped. Context pollution and cost risk went up by the same amount. Reliable use of long context did not arrive with the capacity to hold it.
CLIs, IDE agents, cloud agents and SDKs are becoming one system with several front doors. The churn is real: Google announced on May 19, 2026 that the standalone Gemini CLI was moving to Antigravity CLI, and consumer, Pro and Ultra authentication stopped serving on June 18. Standard and Enterprise licences and paid API‑key use kept working. Roughly a month of notice, and narrower than a full retirement.
The best open models are frequently enormous, and a badly chosen quantization, tool parser, inference server or concurrency policy will destroy the agentic quality the checkpoint advertises.
Provenance
Three research passes disagreed on specifics. Rather than average them into false confidence, here is what held up and what did not.
Three independent passes: direct web research, a codex exec run with web search enabled, and a Gemini pass. Where they agreed, the claim is stated plainly. Where they conflicted, the conflict is shown above rather than resolved by fiat.
The Gemini pass could not run through the Gemini CLI at all. It failed with IneligibleTierError, which is the deprecation described above, so it ran through the Gemini REST API instead. That pass then also produced the weakest numbers of the three, including a SWE‑bench figure and a generation of open‑weight model names that the other two passes contradicted. Its Antigravity claim, though, was exactly right, and the live auth failure confirmed it.
A fourth pass then reviewed the finished page adversarially, and found real errors that all three original passes had shared. The worst is the retracted row above: three sources agreed on a SWE‑bench number because they were all reading the same downstream leaderboard aggregators, none of which had caught up with the benchmark being abandoned. Agreement between models is not verification when they share a source.
The lesson generalises past this page: a single model’s research output is not evidence, and neither is consensus among models drinking from the same well. Independent adversarial review is the verification layer, which is exactly the argument the architecture section makes about never letting the producing model judge its own work.