On this page · 12 sections
- What changed since our earlier comparison
- Claude Opus 5.5: the top independent coding score, at $4/$20
- GPT-6.1 Sol: half the price and the lowest cost per task
- Gemini 4 Argon: $4/$20 from 1 December, with a 1M-token output ceiling
- Head-to-head: prices, caching, batch and limits
- Coding benchmarks, and who ran them
- Which LLM should you pick for coding agents?
- Worked example: monthly cost for a 20-engineer team
- India, EU and US: data residency and procurement
- Rollout checklist
- How eCorpIT can help
- References
Summary. For coding agents going into 2027, the three frontier APIs sit in two price bands. Claude Opus 5.5 (released 22 September 2026) lists at $4 per million input tokens and $20 per million output tokens, with cache reads at $0.20. Gemini 4 Argon sells at an introductory $2/$10 through 30 November 2026 and moves to $4/$20 on 1 December 2026, the same as Opus. GPT-6.1 Sol stays at $2/$10 with cached input at $0.10. All three accept about 1 million tokens of context. On the independent Vals AI Terminal-Bench 4.0 leaderboard (29 October 2026), Opus 5.5 scores 65.15%, Argon 57.58% and GPT-6.1 Sol 55.05%, but Sol's cost per task was $1.72 against $13.20 for Opus and $17.64 for Argon. In our worked model for a 20-engineer team, one identical token profile costs about $2,900 a month on GPT-6.1 Sol, $5,300 on Argon and $5,800 on Opus 5.5. Pick GPT-6.1 Sol for high-volume agent loops, Opus 5.5 for multi-hour tasks where a failed run is expensive, and Argon if you build on Google Cloud or need its 1M-token output ceiling.
About this article. This is for engineering leads and CTOs choosing a model API for coding agents in 2027. We checked every price, limit and date against the vendors' pricing pages, model pages and launch posts, checked benchmark figures against Vals AI's published leaderboards and the vendors' own tables, and recomputed every cost figure, on 1 November 2026. Vendors change prices often, so re-check before you commit; this is not financial advice, and no vendor paid for inclusion. Researched with AI assistance and reviewed by our editorial team.
What changed since our earlier comparison
Earlier this year we compared Claude Opus 5 and GPT-5.6 Sol for coding agents. Almost every number in that piece has since moved.
- Anthropic replaced Opus 5 ($5/$25, cache reads $0.50) with Opus 5.5 at $4/$20 and $0.20 cache reads on 22 September, and says typical workloads cost 40% less than on Opus 5.
- OpenAI shipped GPT-6.1 Sol at DevDay on 29 September at $2/$10, a week after GPT-6 Sol, according to The Batch. The flagship GPT-6 Astra is $10/$50. GPT-5.6 Sol now shows a promotional $4/$20, a cut of 20% on input and 33% on output from $5/$30, available at least through 21 November 2026.
- Google announced Gemini 4 Argon on 30 September and opened it to paid Gemini API customers on 13 October at an introductory $2/$10. Its pricing page dates the switch to $4/$20, with cached input rising from $0.10 to $0.20, on 1 December 2026. Gemini 4 Carbon, announced on 28 October at the same introductory price, is so far limited to a subset of cyber defenders.
- The benchmarks changed too. OpenAI stopped reporting SWE-bench Verified on 23 February 2026 after finding flawed tests and training-data contamination, and Terminal-Bench 4.0 is a fresh set of 66 tasks, so older scores do not carry over.
After 1 December, Opus 5.5 and Argon cost the same per token and GPT-6.1 Sol costs half, but the three use very different numbers of tokens per job.
Claude Opus 5.5: the top independent coding score, at $4/$20
Opus 5.5 (API ID claude-opus-5-5) is the model Anthropic tells most API users to start with, pitched at long-running agentic coding. Its 1M-token context window bills at the standard rate throughout, maximum output is 128K tokens (300K on the Batch API with a beta header), and the reliable knowledge cutoff is June 2026. Adaptive thinking is always on and cannot be switched off.
Pricing. $4 input and $20 output per million tokens. Five-minute cache writes cost $5, one-hour writes $8, and cache hits $0.20 (5% of input). Batch halves the base rates to $2/$10, fast mode costs $8/$40, and US-only inference adds 10%.
Benchmarks. Anthropic reports 66.4% on Terminal-Bench 4.0 at xhigh effort, 54.4% on FrontierCode v1.1 and 57.8% on CursorBench 4.0, all vendor-run. Vals AI's independent run puts Opus 5.5 first of 59 models on Terminal-Bench 4.0 at 65.15%. In its 9 October issue, DeepLearning.AI's The Batch reported Opus 5.5 at the top of the Artificial Analysis Intelligence Index (58), with Sonnet 5.5 and Opus 5.5 in Claude Code holding the top two places on its Coding Agent Index. Google says the index has since moved to version 5.0, with Gemini 4 Carbon first at 59, so treat that ranking as a snapshot.
Limits. Anthropic's safeguards reroute most cybersecurity tasks to Opus 4.8. At launch, Vals reported that counting fallback-assisted tasks as failures would cut its Terminal-Bench 4.0 score from 61.62% to 53.54%. Opus 5.5 is also the heaviest token user of the three on Vals' run: 28.27 million tokens per task against 2.86 million for GPT-6.1 Sol.
Best for: codebase-wide migrations, cross-repository refactors and unattended multi-hour runs. Claude Sonnet 5.5 ($2/$10) scored 64.14% on the same leaderboard but used 58.53 million tokens and $16.51 per task, more than Opus. A cheaper per-token model is not automatically a cheaper agent.
GPT-6.1 Sol: half the price and the lowest cost per task
GPT-6.1 Sol (API ID gpt-6.1-sol) is OpenAI's mid-tier model, pitched as near-Astra performance at a lower cost. It has a 1,050,000-token context window, 128,000 maximum output tokens and an April 2026 knowledge cutoff, and tool calling requires the Responses API.
Pricing. For prompts up to 272K tokens: $2 input, $10 output, $0.10 cached input, and cache writes at $2.50 (30-minute) or $5 (12-hour). Larger prompts pay 2x input and cache rates and 1.5x output for the whole request. Batch and Flex are 50% off ($1/$5), Fast mode is $4/$20, Ultrafast $12/$60, and regional processing adds 10%.
Benchmarks. OpenAI's claim that it nearly matches GPT-6 Astra at one-fifth of Astra's per-token price is vendor-made. Independently, Vals scores it 55.05% on Terminal-Bench 4.0 at $1.72 per task. The Batch reported that Artificial Analysis scored it 52 on its Intelligence Index at max reasoning, one point below Astra, at $0.72 per task against $3.26, and that at xhigh in Codex it beat Astra by one point on the Coding Agent Index at under 15% of Astra's cost per task.
Limits. It has the lowest Terminal-Bench 4.0 score of the three and a surcharge above 272K tokens. The Batch reports that OpenAI classes it as having critical cybersecurity capability, so flagged security requests may go to a fallback model.
Best for: high-volume agent loops such as pull-request review, CI failure fixing and test generation, and teams already on Codex.
Gemini 4 Argon: $4/$20 from 1 December, with a 1M-token output ceiling
Gemini 4 Argon (API ID gemini-4-argon) is Google's flagship reasoning model, aimed at complex software engineering. Its 1,048,576-token window is shared between input and output; default output is capped at 65,536 tokens, but long decode lifts it to 1,048,576. It accepts text, image, video, audio and PDF. Google gives a March 2026 knowledge cutoff, and the model is not on the free tier.
Pricing. $2/$10 through 30 November 2026, then $4/$20, with thinking tokens billed as output. Cached input moves from $0.10 to $0.20, and caching is implicit only, with no explicit cache control and no write fee. Batch and Flex rise from $1/$5 to $2/$10, Priority from $3.60/$18 to $7.20/$36. An invite-only Contributor tier charges $1/$5 but lets Google use the data to improve its products.
Benchmarks. Google's model card reports 77.9% on DeepSWE v1.1 (listing Opus 5.5 at 74.2%) and 57.4% on Terminal-Bench 4.0 (listing Opus 5.5 at 66.4%), both vendor-run. Vals independently scores Argon 57.58% on Terminal-Bench 4.0 at $17.64 per task, and 91.91% on Vibe Code Bench v1.1, second of 123 models.
Limits. The price doubles on 1 December, implicit caching gives you less control over hit rates, and Argon has the highest cost per Terminal-Bench task of the three on Vals' run. Gemini 4 Carbon scores 63.13% on the same leaderboard at $7.45 per task but is not yet open to developers.
Best for: teams on Google Cloud or Antigravity, multimodal work with screenshots, video or long PDFs, and jobs that need one very long output.
Head-to-head: prices, caching, batch and limits
The table draws on each vendor's pricing and model pages as of 1 November 2026.
| Claude Opus 5.5 | GPT-6.1 Sol | Gemini 4 Argon | |
|---|---|---|---|
| API model ID | claude-opus-5-5 | gpt-6.1-sol | gemini-4-argon |
| Released | 22 Sep 2026 | 29 Sep 2026 | Paid API on 13 Oct 2026 |
| Input / output per 1M tokens | $4 / $20 | $2 / $10 (prompts up to 272K) | $2 / $10 to 30 Nov; $4 / $20 from 1 Dec 2026 |
| Cached input (read) | $0.20 | $0.10 | $0.10 to 30 Nov; $0.20 from 1 Dec (implicit only) |
| Cache write | $5 (5-min), $8 (1-hour) | $2.50 (30-min), $5 (12-hour) | No write fee; explicit caching not supported |
| Batch | 50%: $2 / $10 | Batch and Flex 50%: $1 / $5 | Batch and Flex $2 / $10 from 1 Dec |
| Faster tier | Fast mode $8 / $40 | Fast $4 / $20; Ultrafast $12 / $60 | Priority $7.20 / $36 from 1 Dec |
| Long-context pricing | Flat to 1M | 2x input, 1.5x output above 272K | No separate tier listed |
| Context / max output | 1M / 128K (300K on Batch, beta) | 1.05M / 128K | 1,048,576 shared / 65,536 (1M with long decode) |
| Knowledge cutoff | June 2026 | April 2026 | March 2026 |
| Residency premium | US-only inference 1.1x | Regional processing +10% | Not listed on the Gemini API price page |
Coding benchmarks, and who ran them
Vals figures come from its Terminal-Bench 4.0 leaderboard of 29 October 2026 and its model pages; vendor figures come from each launch post or model card.
| Benchmark (who ran it) | Opus 5.5 | GPT-6.1 Sol | Gemini 4 Argon |
|---|---|---|---|
| Terminal-Bench 4.0, Vals AI (independent, mini-SWE-agent, mean of 3 runs) | 65.15% | 55.05% | 57.58% |
| Cost per Terminal-Bench 4.0 task, Vals AI | $13.20 | $1.72 | $17.64 |
| Tokens per Terminal-Bench 4.0 task, Vals AI | 28.27M | 2.86M | 16.66M |
| Terminal-Bench 4.0, vendor-reported | 66.4% (Anthropic, xhigh) | No OpenAI figure checked | 57.4% (Google model card) |
| SWE-bench Multimodal, Vals AI | 70.21% | 43.54% | 56.04% |
| Vibe Code Bench v1.1, Vals AI | 90.29% | 88.93% | 91.91% |
| Code Migration v1.1, Vals AI | 81.01% | 77.32% | 70.80% |
| DeepSWE v1.1, Google model card (vendor-run) | 74.2% | Not in Google's table | 77.9% |
| Artificial Analysis Intelligence Index (as reported 9 Oct 2026) | 58 | 52 | Not in the cited source |
| SWE-bench Verified | Not in Anthropic's launch table | OpenAI stopped reporting it, Feb 2026 | Not in Google's model card |
Three caveats. First, harness and effort matter: Vals runs every model in mini-SWE-agent with bash as the only tool, Opus 5.5 and GPT-6.1 Sol at max effort and Argon at high, while your engineers will use Claude Code, Codex or Antigravity. Second, Terminal-Bench 4.0 is broad: Vals says roughly three quarters of its tasks sit outside traditional software engineering, in areas such as science, ML systems, operations and hardware. Third, the cost figures reflect each model's token habits on long tasks. Dividing Vals' cost per task by its pass rate (our arithmetic) gives roughly $3.12 per solved task for GPT-6.1 Sol, $20.26 for Opus 5.5 and $30.64 for Argon.
Which LLM should you pick for coding agents?
| Pick | If this describes your team |
|---|---|
| GPT-6.1 Sol | You run many short and medium agent loops (PR review, CI fixes, triage), cost per task matters more than the last ten Terminal-Bench points, and prompts stay under 272K tokens |
| Claude Opus 5.5 | Tasks run for hours across large codebases, a failed run costs senior review time, and you can absorb about 2x the bill or route easy work to Sonnet 5.5 |
| Gemini 4 Argon | You are on Google Cloud or Antigravity, need video or long-PDF inputs or outputs beyond 128K tokens, and have priced it at $4/$20, not the November rate |
| A routed mix | You want a flagship on the hardest 20 to 30% of tasks and a cheaper tier for the rest, behind one gateway |
| Wait and re-test | You are Google-first: Gemini 4 Carbon beats Argon on Vals' Terminal-Bench run but is not yet open to developers |
Worked example: monthly cost for a 20-engineer team
This model holds the token profile constant so price differences show; measure your own before you budget.
Assumptions
- 20 engineers, each running 200 agent tasks a month (about 10 per working day), so 4,000 tasks for the team.
- 25 model calls per task, a 50,000-token average prompt (repository context, tool results and history) and 1,200 output tokens per call, including reasoning.
- That is 250 million input and 6 million output tokens per engineer per month, or 5,000 million input and 120 million output tokens for the team.
- 90% of input tokens come from the prompt cache. The other 10% are billed at each vendor's cache-write rate where one is published (Anthropic five-minute $5, OpenAI 30-minute $2.50). Argon has implicit caching only, so its uncached input bills at the normal $4 rate.
- Standard tier, no batch, prompts below 272K tokens, no residency uplift, and the list prices in force from 1 December 2026.
| Team line item, per month | Claude Opus 5.5 | GPT-6.1 Sol | Gemini 4 Argon (from 1 Dec) |
|---|---|---|---|
| Cache reads, 4,500M tokens | 4,500 × $0.20 = $900 | 4,500 × $0.10 = $450 | 4,500 × $0.20 = $900 |
| Uncached input, 500M tokens | 500 × $5.00 = $2,500 | 500 × $2.50 = $1,250 | 500 × $4.00 = $2,000 |
| Output, 120M tokens | 120 × $20 = $2,400 | 120 × $10 = $1,200 | 120 × $20 = $2,400 |
| Monthly total | $5,800 | $2,900 | $5,300 |
| Per engineer per month | $290 | $145 | $265 |
| Per agent task (÷ 4,000) | $1.45 | $0.73 | $1.33 |
| Annual run rate | $69,600 | $34,800 | $63,600 |
At Argon's introductory price the same month costs $2,650 (4,500 × $0.10 + 500 × $2 + 120 × $10), so a pilot run in October or November will look half as expensive as its December invoice.
Sensitivity. Two inputs move these totals more than the choice of vendor.
| Scenario (team, per month) | Opus 5.5 | GPT-6.1 Sol | Argon |
|---|---|---|---|
| Base case, 90% cache hits | $5,800 | $2,900 | $5,300 |
| Cache hit rate falls to 70% | $10,600 | $5,300 | $9,100 |
| Output doubles at higher effort | $8,200 | $4,100 | $7,700 |
| US-only or regional processing (+10%) | $6,380 | $3,190 | Not listed |
Argon is already $500 a month cheaper than Opus 5.5 in the base case because it charges no cache-write premium, and at a 70% hit rate the gap widens to $1,500. A fall from 90% to 70% hits raises every total by roughly 70 to 85%. The equal-token assumption also flatters Opus and Argon: on Vals' run Opus used about ten times GPT-6.1 Sol's tokens per task, and Argon about six times. Against that, Sol solved about ten fewer tasks in every hundred, and a failed run that lands on a senior engineer's desk costs more than the tokens.
Routing. Most teams send only the hardest work to a flagship. With 30% of tasks on the flagship and 70% on a cheaper tier, the same token profile and current list prices:
| Routed mix (30% / 70%) | Cheaper tier list price | Cheaper tier on Vals Terminal-Bench 4.0 | Team cost per month |
|---|---|---|---|
| Opus 5.5 + Sonnet 5.5 | $2 / $10, cache reads $0.10 | 64.14% | $3,770 |
| GPT-6.1 Sol + GPT-6.1 Terra | $1 / $5, cached input $0.05 | 49.49% | $1,885 |
| Argon + Gemini 3.8 Flash (from 1 Jan 2027) | $1.50 / $7.50, cached input $0.15 | 19.70% | $3,218 (excluding cache storage) |
The cheaper tier alone would cost $2,900 on Sonnet 5.5, $1,450 on GPT-6.1 Terra and $2,325 on Gemini 3.8 Flash; each mix is 0.3 × the flagship total plus 0.7 × that figure. Google's cheaper tier is the weak link: Gemini 3.8 Flash doubles in price on 1 January 2027, and on Vals' run it scored far lower and cost $8.65 per task, five times GPT-6.1 Sol. Our comparison of Gemini 3.8 Flash, Claude Haiku 5.5 and GPT-6 Luna covers the budget tier in detail.
Seats versus API. Claude Team and ChatGPT Business premium seats both cost $100 per seat per month on annual billing ($125 monthly), with usage limits, and both include the vendor's coding agent, Claude Code or Codex. Twenty premium seats cost $2,000 a month, less than any API total above, but seat limits cap what one engineer can run, and ChatGPT and Codex will share one usage limit from December. Claude Enterprise is $20 per seat (20-seat minimum) plus usage at API rates, so this team pays $400 before tokens. Headless CI agents need API keys either way; a ChatGPT Business subscription does not include API usage. Large reported bills, such as those in Microsoft and Uber's Claude Code token bills, came from usage volume rather than the rate card.
India, EU and US: data residency and procurement
- Anthropic: the
inference_geoparameter accepts onlyus(at 1.1x) orglobal, andusis the only workspace geo for stored data. Opus 5.5 is available with zero data retention and, per Anthropic, carries watermarking to comply with the EU AI Act.
- OpenAI: the India endpoint (
in.api.openai.com) offers regional storage but not regional processing, and requires Modified Abuse Monitoring or Zero Data Retention approval. Regional processing adds 10% for models released on or after 5 March 2026, and FedRAMP endpoints add 10% from 1 November 2026.
- Google: we found no India regional processing option for Argon on the Gemini API pages we checked. Paid Gemini API traffic is not used to improve Google's products; the Contributor tier is the exception.
- DPDP Act 2023: fixtures, log samples and support tickets pasted into agent sessions often contain personal data. Redact before the prompt and run agents in a controlled sandbox; our coding agent sandboxes comparison covers the isolation options.
- Billing: all three list API prices in US dollars, so rupee budgets carry exchange-rate exposure.
Rollout checklist
- Log one normal week of agent traffic and record tokens, cache hits and tasks per engineer.
- Run a two-week bake-off on 30 to 50 of your own tasks, same harness for all three models; record pass rate, tokens and cost per merged pull request.
- Put stable content first in every prompt (system prompt, tool definitions, repository map) so caches stay warm, especially on Argon.
- Route through an AI gateway with per-team budgets so you can switch models without touching engineers' tooling.
- Set a hard monthly cap per engineer and alert at 80%.
- Move non-interactive jobs such as nightly review and bulk test generation to Batch or Flex at 50% off.
- Diary 21 November (the earliest end of the GPT-5.6 Sol promotion), 1 December (Argon's price switch and ChatGPT's shared ChatGPT and Codex limit), 1 January 2027 (Gemini 3.8 Flash doubles) and Gemini 4 Carbon's developer release.
How eCorpIT can help
eCorpIT is a Gurugram-based technology organisation, founded in 2021, assessed at CMMI Level 5 and MSME certified, with senior-led engineering teams working across AWS, Microsoft and Google platforms. Our AI coding agent rollout and governance service runs a model bake-off on your own repositories, instruments tokens and cache hits, and sets up gateway routing with per-team budgets. Talk to us at /contact-us/.
Last updated: 1 November 2026.
