Straitly compass logostraitly

CODING AGENT COST · CALCULATOR

Cheapest Claude API for Claude Code and Cursor

The cheapest Claude API for Claude Code and Cursor is one that bills the same models below list price: Straitly bills Claude, GPT, Gemini and Grok about 40% below list, with no fee on top.[1] At list price, Claude Code averages about $13 per developer per active day.[2]

Last updated . Every number links to its source.

One developer, one month on Claude Opus 5.5Anthropic's 996.5K-token sample session, 10 a day for 21 days

  • $855

    No cache, list priceEvery input token at full price.[14]

  • $115

    With a cache, list priceMost tokens are cheap cache reads.[2][15]

  • $69

    With a cache, on StraitlySame session, no fee on top.[23]

What does one Claude Code session cost on each model?

Anthropic's sample session used 996.5K tokens and cost $0.55 on Claude Sonnet 4.6. Claude Code resends the conversation with every request, so 94% of those tokens were cache reads.[2]

  • Cache reads 94.3%
  • Cache writes 5.0%
  • Output 0.5%
  • Uncached input 0.1%
“With prompt caching, Claude Code re-reads that history at the cached token rate, so a one-line question in a session that has been open all day still draws usage for the whole conversation.”
[2] Claude Code docs, Checked 9 November 2026
Sample session repriced, official vs Straitly
MODELPER SESSIONPER MONTHSAVING
Claude Fable 5.1read $0.25 / $0.15, out $50.00 / $30.00Per session$1.14 / $0.682Per month$239 / $143Saving40%
Claude Opus 5.5read $0.20 / $0.12, out $20.00 / $12.00Per session$0.549 / $0.329Per month$115 / $69.15Saving40%
Claude Sonnet 5.5read $0.10 / $0.12, out $10.00 / $6.00Per session$0.274 / $0.221Per month$57.62 / $46.42Saving19%
GPT-6.1 Solread $0.10 / $0.060, out $10.00 / $6.00Per session$0.274 / $0.165Per month$57.62 / $34.57Saving40%
Gemini 4 Argonread $0.10 / $0.060, out $10.00 / $6.00Per session$0.249 / $0.15Per month$52.37 / $31.42Saving40%
DeepSeek V4.1 Flashread $0.0030 / $0.0018, out $0.60 / $0.36Per session$0.014 / $0.008Per month$2.87 / $1.72Saving40%

Official / Straitly. A month is 210 sessions.[14][16][17][18][23]

How the $0.55 adds up, and cost per million agent tokens

(1.2K × $3.00 + 5.3K × $15.00 + 940K × $0.30 + 50K × $3.75) ÷ 1M = $0.553

Sonnet 4.6 list price per million: $3.00 input, $15.00 output, $0.30 cache read, $3.75 cache write. Cache hit rate on input: 94.8%.[2][14] A month is 10 sessions a day for 21 working days. Cache writes are priced at the short-lived rate, 1.25x input at Anthropic and OpenAI; Google and DeepSeek charge no write fee. With no cache, Claude Opus 5.5 costs $4.08 per million agent tokens instead of $0.54, so caching cuts it 87%.

Cost per million agent tokens

Official list price at a 95% cache hit rate and 0.5% output, close to Anthropic's sample session.

  1. GPT-6 AstraStraitly $1.09$1.82
  2. Claude Fable 5.1Straitly $0.66$1.11
  3. Claude Opus 5.5Straitly $0.32$0.54
  4. Claude Sonnet 5.5Straitly $0.22$0.27
  5. GPT-6.1 SolStraitly $0.16$0.27
  6. Gemini 4 ArgonStraitly $0.15$0.24
  7. GPT-6.1 TerraStraitly $0.081$0.13
  8. GPT-6 LunaStraitly $0.011$0.018
  9. DeepSeek V4.1 FlashStraitly $0.0080$0.013

Bar lengths use a log scale so small and large values fit on one chart.

[14] Anthropic, Checked 9 November 2026; [16] OpenAI, Checked 9 November 2026; [17] Google, Checked 9 November 2026; [18] DeepSeek API docs, Checked 9 November 2026; [23] Straitly, Snapshot 9 November 2026

“Thinking tokens are billed as output tokens, and the default budget can be tens of thousands of tokens per request depending on the model.”
[2] Claude Code docs, Checked 9 November 2026

How much does Claude Code cost per developer?

Agents resend their whole context on every turn, so they use many times the tokens of chat.

  • $13

    Average Claude Code cost per developer per active day across enterprise deployments, or $150-250 per developer per month.[2] Claude Code docs, Checked 9 November 2026

  • 90%

    Share of Claude Code users whose costs stay below $30 per active day.[2] Claude Code docs, Checked 9 November 2026

“agentic tasks are uniquely expensive, consuming 1000x more tokens than code reasoning and code chat, with input tokens rather than output tokens driving the overall cost”
[4] Bai et al., 24 Apr 2026, revised 2 Oct 2026

How many more tokens agents use

Each against its own baseline.

  1. Anthropic's agents vs chat4x
  2. Agent team in plan mode vs one session~7x
  3. Multi-agent systems vs chat15x
  4. Same task, most vs fewest tokens across runs30x
  5. Agentic coding vs code chat1000x

Bar lengths use a log scale so small and large values fit on one chart.

[3] Anthropic Engineering, 13 Jun 2025; [2] Claude Code docs, Checked 9 November 2026; [4] Bai et al., 24 Apr 2026, revised 2 Oct 2026

All 11 more stats in this section
How much does Claude Code cost per developer?
$0.55Total cost in Anthropic's sample /usage readout: 1.2K input, 5.3K output, 940K cache read and 50K cache write tokens on Claude Sonnet 4.6.[2] Claude Code docs, Checked 9 November 2026
94%Share of all tokens in that sample session that were cache reads. Claude Code resends the full conversation with every request.[2] Claude Code docs, Checked 9 November 2026
~7xExtra tokens agent teams use against a standard session when teammates run in plan mode, because each teammate keeps its own context window.[2] Claude Code docs, Checked 9 November 2026
10KsDefault extended-thinking budget per request on some models: tens of thousands of tokens, billed as output.[2] Claude Code docs, Checked 9 November 2026
4x / 15xTokens Anthropic's agents use against chat, and multi-agent systems against chat, in Anthropic's own data.[3] Anthropic Engineering, 13 Jun 2025
1000xMore tokens consumed by agentic coding tasks than by code reasoning and code chat, across eight frontier models on SWE-bench Verified. Input tokens drive the cost.[4] Bai et al., 24 Apr 2026, revised 2 Oct 2026
30xSpread in total tokens between runs of the same model on the same task. Higher token use did not buy higher accuracy.[4] Bai et al., 24 Apr 2026, revised 2 Oct 2026
1.5M+Extra tokens per task Kimi-K2 and Claude Sonnet 4.5 used on average against GPT-5 on the same tasks.[4] Bai et al., 24 Apr 2026, revised 2 Oct 2026
0.39Best correlation between a model's own prediction of its token use and the real figure. Models systematically underestimate.[4] Bai et al., 24 Apr 2026, revised 2 Oct 2026
79-98%Coding tasks affected by three wasteful habits (re-reading code, rewriting similar scripts, re-running tests), worth up to 22.75% of task cost.[5] Hu et al., 25 Sep 2026, revised 12 Oct 2026
QuadraticHow token use grows with each model call, because the whole conversation history is passed back in as input every time.[7] Epoch AI, Checked 9 November 2026

What does one coding task cost?

Vals AI ran each model on SWE-bench Verified, 500 real GitHub issues. Models within 2 points of each other can differ 10x or more in cost.[6]

  • $0.47

    Median cost per task across 80 runs, from $0.01 to $4.00.[6] Vals AI, Results as of 1 Sep 2026

  • 1.34M

    Median tokens per task across 78 runs, from 81.3K to 5.92M.[6] Vals AI, Results as of 1 Sep 2026

  • 2.8x

    gpt-5 at high reasoning effort against low on Aider's benchmark, for 6.7 more points.[8] Aider, Runs dated Dec 2024 to Oct 2025

SWE-bench Verified cost and tokens per task, Vals AI
MODELSCORECOST / TASKTOKENS / TASKON STRAITLY
Claude Opus 5Score97.0%Cost$1.29Tokens1.09MStraitly$0.774
DeepSeek V4 Pro 0813Score96.4%Cost$0.10Tokens764KStraitly-
GPT-5.6 SolScore96.2%Cost$1.15Tokens512.4KStraitly$0.489
Grok 4.6Score95.6%Cost$0.78Tokens969KStraitly-
GPT-5.6 TerraScore95.4%Cost$0.40Tokens618.9KStraitly-
GLM 5.3Score95.4%Cost$0.34Tokens958.2KStraitly$0.091-$0.107
Claude Fable 5Score95.0%Cost$2.05Tokens722.6KStraitly$1.23
Kimi K3Score93.4%Cost$0.76Tokens884KStraitly$0.388
GPT-5.6 LunaScore93.0%Cost$0.04Tokens781KStraitly-
GLM 5.3 FlashScore92.0%Cost$0.02Tokens792.8KStraitly$0.005
Gemini 3.8 FlashScore80.0%Cost$2.19Tokens5.92MStraitly$1.31
Claude Sonnet 5Score79.6%Cost$1.49Tokens2.94MStraitly$0.894
Show all 12 rows

Mini-SWE-agent harness, sorted by score. [6] Vals AI, Results as of 1 Sep 2026

Aider effort costs, and how to read the task table
  • On Straitly (est.) = Vals cost × (Straitly price ÷ official price). It is a range across the input, output and cache-read ratios, because Vals does not publish the token split. Straitly prices are from the catalog snapshot of 9 November 2026; a dash means the model is not in the index.[23]
  • Inside the vendors' own agents: Claude Opus 4.8 in Claude Code scored 85.8% at $0.67 a task; GPT 5.5 in Codex scored 76.4% at $0.65 a task; GPT 5.5 in Factory scored 76.2% at $1.13 a task.[6]
  • Cost bars use a log scale so $0.02 and $2.19 fit in one column. Token bars are linear.
  • Aider publishes the dollar cost of a full run of its polyglot benchmark, 225coding exercises, at the model's API price on the run date. Thinking tokens are billed as output.[8]
Aider polyglot cost per exercise
MODELPASSRUN COSTPER EXERCISE
gpt-5 (low effort)25 Aug 202581.3%$10.37$0.046$10.37 ÷ 225
gpt-5 (medium effort)25 Aug 202586.7%$17.69$0.079$17.69 ÷ 225
gpt-5 (high effort)23 Aug 202588.0%$29.08$0.129$29.08 ÷ 225
o3-pro (high)28 Jun 202584.9%$146$0.65$146 ÷ 225
claude-opus-4 (32k thinking)25 May 202572.0%$65.75$0.292$65.75 ÷ 225
DeepSeek-V3.2-Exp (Reasoner)3 Oct 202574.2%$1.30$0.006$1.30 ÷ 225
DeepSeek-V3.2-Exp (Chat)3 Oct 202570.2%$0.88$0.004$0.88 ÷ 225
o1 (high)21 Dec 202461.7%$187$0.833$187 ÷ 224

What would your coding agent cost?

Defaults follow Anthropic's sample session. Straitly prices as of 9 November 2026.

Your usage

Official $4.00 in, $20.00 out, $0.20 cache read per 1M. Straitly $2.40 in, $12.00 out, $0.12 cache read.

Anthropic's sample session is about 1,000K; the median SWE-bench task on Vals AI about 1,340K.

95%
0.5%

OFFICIAL / MONTH

$67.76

STRAITLY / MONTH

$40.66

Official, no cache
Official list price
On Straitly

Dashed lines: $20, $100 and $200 flat-rate plans.

Per session, official
$0.54
Per session, Straitly
$0.32
Same month with no cache
$514
Saved vs official per month
$27.11

At this usage a $100 flat-rate plan costs more than paying per token on Straitly, unless you need its extras.

Uncached input is billed as a cache write (1.25x input).

About the models in the calculator
  • Claude Haiku 5.5 is left out because its price changes above 100K-token prompts, which agent sessions often pass.
  • Gemini 4 Argon's input price doubles on 1 December 2026.
  • DeepSeek is shown at its off-peak rate; peak rates are double.
  • Input not served from cache is billed as a cache write, because agents write new context to the cache. Thinking tokens count as output. List prices for standard requests; no batch or fast mode.

Is Claude Max or ChatGPT Pro cheaper than the API?

For heavy users who stay inside a plan's limits, often yes; below break-even, paying per token costs less. Cursor bills Claude, GPT and Gemini at API list price on every plan.[11]

  • 372M

    Claude Opus 5.5 agent tokens $200 buys at list price, about 18 sample sessions a day.[14] Anthropic, Checked 9 November 2026

  • 620M

    The same $200 on Straitly, about 30 sessions a day.[23] Straitly, Snapshot 9 November 2026

Flat-rate coding agent plans
PLANPRICEWHAT IT INCLUDES
Claude Pro, $20/mo ($17 billed yearly)$20/mo ($17 billed yearly)Usage limits every 5 hours and every week
Claude Max, $100/mo$100/mo5x Pro usage, plus $100 a month in API credits
Claude Ultra, $200/mo$200/mo10x Pro usage a week, plus $200 a month in API credits
ChatGPT Plus, $20/mo$20/mo15-160 local GPT-6.1 Sol messages per 5 hours (estimate)
ChatGPT Pro, $100, $200 or $500/mo$100, $200 or $500/moNo 5-hour limit; weekly limits may apply
Cursor Premium, $30/mo$30/mo4x Plus usage of Cursor's own models (Grok, Composer); Claude, GPT and Gemini at API list price
Cursor Super, $100/mo$100/mo15x Plus usage of Cursor's own models; third-party models at API list price
Cursor Ultra, $200/mo$200/mo50x Plus usage of Cursor's own models; third-party models at API list price
Show all 8 rows

Each vendor's pricing page. [9] Anthropic, Checked 9 November 2026; [10] OpenAI, Checked 9 November 2026; [11] Cursor docs, Checked 9 November 2026

How many tokens each plan price is worth, by model

Plan price ÷ cost per million agent tokens (95% cache hits, 0.5% output). Each cell: monthly tokens at official list / on Straitly, then the 996.5K-token sample sessions a day over 21working days that equals. Use more than this inside the plan's limits and the plan wins; use less and paying per token wins.

Monthly tokens that equal a plan price
MODEL$20 PLAN$100 PLAN$200 PLAN
Claude Opus 5.5$0.54 / $0.32 per 1M (official / Straitly)$20 plan37M / 62M~2 / 3 sessions a day$100 plan186M / 310M~9 / 15 sessions a day$200 plan372M / 620M~18 / 30 sessions a day
Claude Sonnet 5.5$0.27 / $0.22 per 1M (official / Straitly)$20 plan74M / 92M~4 / 4 sessions a day$100 plan372M / 459M~18 / 22 sessions a day$200 plan744M / 917M~36 / 44 sessions a day
GPT-6.1 Sol$0.27 / $0.16 per 1M (official / Straitly)$20 plan74M / 124M~4 / 6 sessions a day$100 plan372M / 620M~18 / 30 sessions a day$200 plan744M / 1,240M~36 / 59 sessions a day
Gemini 4 Argon$0.24 / $0.15 per 1M (official / Straitly)$20 plan82M / 137M~4 / 7 sessions a day$100 plan410M / 683M~20 / 33 sessions a day$200 plan820M / 1,366M~39 / 65 sessions a day
  • Plans cap usage per 5 hours or per week, so the token figure is a ceiling on value, not a promise. OpenAI says API token prices are separate from subscription usage and should not be used to estimate included tasks.[10]
  • A lower per-token price applies to usage you pay per token. Cursor's own models and Tab completion stay billed by Cursor.[12]
  • Cline is free and open source and bills only inference, through its own credits or your API key. Aider is also open source and runs on your own key.[13]

How do you cut Claude Code and Cursor costs?

Caching and a lower token price do the most, and they stack. See caching savings, and the price index.

“These results show that token reduction is not a reliable proxy for cost reduction in tool-heavy coding agents.”
[20] Weinberger and Hozez, 13 Jul 2026
Coding agent cost levers
#LEVEREFFECT
01Keep the cache warmCache reads cost 0.05x-0.1x input. At 95% hits Opus 5.5 falls from $4.08 to $0.54 per million agent tokens (87% less).[15] Anthropic docs, Checked 9 November 2026
02Pay less per tokenClaude, GPT, Gemini and Grok about 40% below list on Straitly (median 40% across 102 models). Stacks with every lever.[22] Straitly, Snapshot 9 November 2026
03Clear or compact context between tasksClaude Code re-reads the whole conversation each turn, so a long-open session costs more per question.[2] Claude Code docs, Checked 9 November 2026
04Lower reasoning effort for routine workThinking is billed as output. On Aider, gpt-5 high cost 2.8x low for 6.7 more points.[8] Aider, Runs dated Dec 2024 to Oct 2025
05Pick the model per taskOn SWE-bench Verified, models within 2 points differ 10x or more in cost per task.[6] Vals AI, Results as of 1 Sep 2026
5 more levers
More coding agent cost levers
#LEVEREFFECT
06Keep agent teams smallTeams use about 7x the tokens of one session in plan mode. Each teammate is billed until it exits.[2] Claude Code docs, Checked 9 November 2026
07Preprocess tool output with hooksFiltering a 10,000-line log to error lines cuts that context from tens of thousands of tokens to hundreds.[2] Claude Code docs, Checked 9 November 2026
08Write project skills for repeated workDeveloper-written skills cut cost by up to 41.73% in a 10K-trajectory study.[5] Hu et al., 25 Sep 2026, revised 12 Oct 2026
09Trim wasted context, carefullyRemoving stale trajectory content cut total cost 21-36%. Aggressive tool-output compression raised billed cost 6.8%.[19] Xiao et al., 28 Sep 2025
10Add an external cost controller18-31% lower token cost across Claude Code, Codex and GitHub Copilot CLI with comparable task success.[21] Li et al., 15 Oct 2026

What do people ask about coding agent cost?

What is the cheapest Claude API for Claude Code and Cursor?

One that bills the same Claude models below Anthropic's list price. Straitly bills Claude, GPT, Gemini and Grok about 40% below list with no fee on top, so Claude Opus 5.5 at a 95% cache hit rate costs about $0.32 per million agent tokens instead of $0.54.

How much does Claude Code cost on API tokens?

Anthropic reports an average of about $13 per developer per active day across enterprise deployments, or $150-250 per developer per month, with 90% of users below $30 per active day. A single sample session in Anthropic's docs used about 1 million tokens and cost $0.55 on Claude Sonnet 4.6.

Is a Claude Max or ChatGPT Pro plan cheaper than the API?

For heavy users, often yes. At Claude Opus 5.5 list price, a coding agent session with a 95% cache hit rate costs about $0.54 per million tokens, so a $200 plan equals about 372M tokens a month. If you use more than that and stay inside the plan's limits, the plan is cheaper. Below it, paying per token costs less.

How do I reduce coding agent costs?

Keep the prompt cache warm (at a 95% hit rate it cuts Claude Opus 5.5 cost per agent token by 87%), pay a lower price per token, clear or compact context between tasks, lower reasoning effort for routine work, choose a cheaper model where it scores close enough, keep agent teams small and filter tool output before the model sees it.

Does Straitly work with Claude Code, Cursor and Codex?

Yes. Each tool accepts a custom base URL and API key, and Straitly bills a median 40% below official list price. Usage inside a flat-rate plan, and Cursor's own models, stay billed by the tool's vendor.

How do you use a cheaper Claude API in Claude Code, Cursor and Codex?

Point the tool at Straitly's base URL with your key. Zero data retention is on by default for every account and key: prompts and completions are never written to disk or logs and never used for training (data retention).

    Sources and method

    How these numbers were checked

    Every figure was read from the source listed on 9 November 2026. Official prices are list prices in USD per million tokens for standard requests. Straitly prices come from the catalog snapshot of 9 November 2026. Per-session and per-month figures are worked examples from published token counts, not measurements of anyone's usage. Benchmark costs are at the API price on each run date.

    1. [1]Straitly pricing Straitly, Checked 9 November 2026
    2. [2]Manage costs effectively Claude Code docs, Checked 9 November 2026
    3. [3]How we built our multi-agent research system Anthropic Engineering, 13 Jun 2025
    4. [4]How Do AI Agents Spend Your Money? Analyzing and Predicting Token Consumption in Agentic Coding Tasks Bai et al., arXiv 2604.22750, 24 Apr 2026, revised 2 Oct 2026
    5. [5]Analyzing and Mitigating Cost-Inefficient Behaviors in Coding Agents Hu et al., arXiv 2609.30725, 25 Sep 2026, revised 12 Oct 2026
    6. [6]SWE-bench Verified leaderboard (archived results) Vals AI, Results as of 1 Sep 2026
    7. [7]SWE-bench Verified: methodology Epoch AI, Checked 9 November 2026
    8. [8]Aider LLM leaderboards (polyglot benchmark) Aider, Runs dated Dec 2024 to Oct 2025
    9. [9]Plans and pricing Anthropic (Claude), Checked 9 November 2026
    10. [10]Codex pricing and usage limits OpenAI, Checked 9 November 2026
    11. [11]Models and pricing Cursor docs, Checked 9 November 2026
    12. [12]Bring your own API key Cursor docs, Checked 9 November 2026
    13. [13]Cline pricing Cline, Checked 9 November 2026
    14. [14]Claude API pricing Anthropic, Checked 9 November 2026
    15. [15]Prompt caching Anthropic docs, Checked 9 November 2026
    16. [16]API pricing OpenAI, Checked 9 November 2026
    17. [17]Gemini Developer API pricing Google, Checked 9 November 2026
    18. [18]Models and pricing DeepSeek API docs, Checked 9 November 2026
    19. [19]Reducing Cost of LLM Agents with Trajectory Reduction Xiao et al., arXiv 2509.23586, 28 Sep 2025
    20. [20]Token Reduction Is Not Cost Reduction Weinberger and Hozez, arXiv 2607.12161, 13 Jul 2026
    21. [21]TokenEconomizer: Experience-Driven Black-Box Cost Control for Coding Agents Li et al., arXiv 2610.22732, 15 Oct 2026
    22. [22]Straitly price snapshot vs official list and OpenRouter (164 models) Straitly, Snapshot 9 November 2026
    23. [23]Straitly live model catalog Straitly, Snapshot 9 November 2026
    Show all 23 sources