Fidwise
Every AI coding tool,
actually compared.
Price, quota, measured quality, privacy and how likely each one is to still exist next year. Checked against the companies' own pages, not blog posts.
40+ options
Every price sourced and dated
Unverified items named
12 August 2026 · Dollar figures converted at 88 rupees
The question
Can we get better than what
we pay for today, for less?
Where we stand
- We pay 880 rupees a month for the coding assistant.
- Normal day: 100 to 300 requests. Heavy shipping day: 1,150.
- Money is tight, so every rupee has to earn its place.
What we refuse to trade
- Quality. A cheaper assistant that writes worse code costs more, in time.
- Ownership. We ship a real commercial product.
- Stability. A tool that dies mid project is worse than no tool.
How this was checked
Read from the source,
not from a blog post.
1
Company pages only
Prices, limits and privacy terms taken from each company's own pricing page, documentation and terms. Every figure carries a source and a date.
2
Rumours thrown out
Several confident numbers circulating online traced back to auto generated content farms. They were discarded rather than repeated.
3
Gaps admitted
Where a page could not be read, it is marked unverified instead of filled with a plausible guess. The last slide lists every gap.
Three separate sweeps. Each one overturned something from the one before.
Finding one
Companies mark
their own homework.
The same model, on the same test, scores wildly differently depending on who ran it.
GLM, as claimed by its makerZhipu, own model card
81
GLM, on the independent boardTerminal-Bench, verified run
58.7
A small open model, neutral testNVIDIA's own harness
15.2
Rule we adopted: only trust a score somebody else measured.
Sources: Zhipu GLM 5.2 model card; Terminal-Bench 2.1 verified leaderboard (tbench.ai), 1 May 2026; NVIDIA Nemotron 3.5 card, 11 Aug 2026.
The honest ranking
Who actually writes
the best code.
Independently verified terminal coding test. Higher is better. The cost column is what one full test run cost the submitter.
| Model | Score | Run cost | Notes |
|---|
| Claude Fable 5 | 83.8 | $553 | Best measured. Also the most expensive to run. |
| GPT-5.5 | 83.1 | $2,059 | Near identical result, four times the cost. |
| Grok 4.5 | 79.3 | $134 | penalised Docked 9 points for gaming the test. |
| Claude Opus 4.8 | 78.9 | $287 | Clean run. |
| GPT-5.6 Luna | 75.7 | $241 | the bargain Cheapest frontier model by far. |
| Claude Sonnet 5 | 74.6 | $288 | Clean run. |
| Gemini 3.1 Pro | 65.8 | $236 | A clear step down. |
| GLM (Chinese, open) | 58.7 | $277 | Best open weight result measured independently. |
Terminal-Bench 2.1 verified leaderboard, tbench.ai, entries dated 1 May to 11 Jul 2026. Every row checked by a Terminal-Bench reviewer. Scores carry roughly plus or minus 1.3 points.
Finding two
Free tiers keep
dying.
Five that were recommended online this year no longer exist. Two of them I recommended myself before checking.
| What it was | What happened | When |
|---|
| Qwen free coding tier | Shut down | April 2026 |
| Google's free Gemini terminal tool | Stopped serving free users | 18 June 2026 |
| GitHub's free model service | Retired completely | 30 July 2026 |
| Cerebras free tier | Replaced by a 30 day trial, card required | 2026 |
| Amazon's free coding tool | Discontinued, successor gives about 90 rupees of credit | 2026 |
Anything free is a marketing budget, and marketing budgets get cut.
Sources: each vendor's own retirement notice or documentation, all read 12 Aug 2026.
What survives
Genuinely free,
as of today.
| Option | What you get | Card? | The catch |
|---|
| Google Antigravity | Claude Opus 4.6 and Sonnet 4.6, plus Google's own models | No | Refills weekly, so a heavy day locks you out for days |
| OpenRouter, free models | 50 requests a day | No | 50 a day is roughly one hour of real work |
| OpenRouter, after one payment | 1,000 requests a day, permanently | Once | One time 880 rupees, then nothing ever again |
| Trae free tier | Unlimited autocomplete, limited agent use | No | Owned by ByteDance. Trains on your chats unless you switch privacy mode on |
| Grok free tier | Grok 4.5 and their coding product | No | Not a terminal tool on its own, needs a separate app to drive it |
| Run it on your own machine | Unlimited, private | No | Needs expensive hardware and still writes worse code. See later slide |
All verified from vendor pricing and documentation pages, 12 Aug 2026.
The best free option
Google is giving away
Anthropic's best model.
- Google's free plan lists Claude Opus 4.6 and Sonnet 4.6, Anthropic's frontier models, at zero cost.
- It includes a real terminal assistant, not just a chat window.
- No credit card, and India is an approved country.
- Google is paying Anthropic's bill to win developers away from competitors.
Verified on two separate Google pages, by hand.
Why this is strange
Claude Opus costs roughly 440 rupees per million words in and 2,200 out if you buy it directly from Anthropic.
Google hands it to you for nothing, then quietly limits how much you can use.
antigravity.google/pricing and /docs/models, read directly 12 Aug 2026.
The fine print
Free, but read
the terms first.
!
Staff can read your code
Their terms say Google employees and contractors may access, view and review what you send, and it is used to improve their systems. There is an off switch, but it starts on.
antigravity.google/terms
!
Accounts get banned
Using it from any other tool breaks their terms and they have banned real accounts for it. Sign in is a personal Google account, so a ban costs more than the tool.
Terms clause 6, plus reported bans
!
Nobody knows the limit
Google publishes no number. Paying customers report running dry and waiting a week. The allowance is measured in work done, not requests, and can change at any time.
antigravity.google/docs/plans
Usable, on a spare account, for work that is not sensitive.
The card already in your pocket
Your phone plan may
fix the one real flaw.
The problem
On the free plan the allowance refills once a week. Burn it on Monday and you wait until the next week.
The fix
On Google's paid plan it refills every five hours. Same tool, same models, a completely different working day.
Why it may cost nothing
- Jio gives eligible customers 18 months of Google's paid plan free, a stated value of about 35,100 rupees.
- Google's own help page lists Antigravity with higher limits as a benefit of that paid plan.
- Claimed inside the MyJio app. No card, no payment.
blog.google India announcement and Google One help page, read 12 Aug 2026.
Not yet confirmed: whether the Jio version is the full paid plan. Its announcement lists older models and less storage than the current paid plan, so the only certain answer is to claim it and look.
Finding three
The Chinese models are
genuinely good now.
And most are released under licences that let anyone use them commercially, free, forever. That is the part that matters more than the score.
| Model | Licence | Can we use it commercially? | Note |
|---|
| DeepSeek V4 | MIT, fully open | Yes, no conditions | Cleanest terms of anyone. They assign the output to you in writing. |
| GLM 5.2 | MIT, fully open | Yes, no conditions | No regional limits. |
| Qwen3 Coder | Apache 2.0 | Yes, no conditions | Licence file read directly to confirm. |
| Kimi K3 | Custom | Yes for us | Conditions only bite above 20 million dollars of revenue. |
| MiniMax M3 | Community licence | No, not by default | Non commercial unless you display their name in the product and email them. |
One licence trap in the set, and it is the one nobody mentions.
Licence files read directly from each model's repository, 12 Aug 2026.
Finding four
Who is allowed to
read your code.
| Provider | Trains on your work? | Can you turn it off? | Evidence |
|---|
| Google Antigravity | Yes, by default | Yes, in settings | Staff review is stated explicitly in the terms |
| Google's free coding interface | Yes | No | Their terms tell you not to send confidential material |
| Claude, personal plans | Only if switched on | Yes | Kept 5 years if on, 30 days if off |
| Claude, pay per use | No | Not needed | Stated in their terms |
| OpenAI, pay per use | No | Not needed | Unless you deliberately opt in |
| DeepSeek | Yes, with a named off switch | Yes, self service | They also assign output ownership to you |
| Trae, by ByteDance | Yes, until you switch privacy mode on | Yes, free | Your project files are never trained on, they say |
| MiniMax | Reserves commercial rights over the data | No stated option | Stored in the United States |
The widely repeated claim that ByteDance keeps your code for five years with no way out could not be found in any official document, and their own current documentation contradicts it. We are not repeating it as fact.
The surprise
What we already pay for
is quietly excellent.
880 rupees buys
- 1,320 rupees of model usage. More value than the price.
- Every top model on the market: Claude Opus 5, Fable 5, GPT-5.6, Gemini, Grok 4.5, Kimi.
- Autocomplete that costs nothing against the allowance.
- The terminal assistant included, no separate charge.
Nobody else does this
No other company checked gives more model usage than the subscription costs.
The catch: the allowance resets on the first of each month and anything unused is lost.
GitHub documentation, read 12 Aug 2026.
We went looking for something cheaper and found we were already on it.
The biggest lever
Which model you pick matters
40 times more than which tool.
Same subscription, same terminal, one dropdown. This is the single highest value change available to us.
| Model | Quality | Cost per million words in | Out | Work per month on our plan |
|---|
| Claude Fable 5 | 83.8 | 880 rupees | 4,400 | A few long sessions, then empty |
| Claude Opus 5 | 78.9 | 440 | 2,200 | Roughly twice as far |
| GPT-5.6 Luna | 75.7 | 18 rupees | 106 | Around 40 times further |
Luna gives about 90 percent of the best result for about 2 percent of the cost.
Quality from the verified Terminal-Bench leaderboard. Prices from GitHub's published per model rates, 12 Aug 2026. A further 10 percent discount applies with automatic model selection.
If we ever want a second one
The cheapest paid options
worth considering.
| Option | Per month | What you get | Worth it? |
|---|
| What we pay now | 880 | 1,320 rupees of usage, every top model | Benchmark to beat |
| Trae Lite | 264 | 440 rupees of usage, unlimited autocomplete | Cheapest paid tier anywhere |
| Trae Pro | 880 | 1,760 rupees of usage | Same price as ours, more usage, worse privacy |
| DeepSeek, pay as you go | roughly 250 | Near frontier quality, you pay only for what you use | Best value per rupee |
| GLM coding plan | 1,584 | Large allowance, half price outside peak hours | Was 264 rupees, raised six fold |
| Claude, Grok or ChatGPT paid | 2,640 and up | One company's models only | Three times our current bill |
All from vendor pricing pages, 12 Aug 2026. Some tier prices could not be read and are excluded rather than guessed.
The tempting dead end
Buying hardware means
paying to get worse.
Spend 2 lakh on a graphics card
- Best model it can actually run scores 51.7.
- It forgets faster, holding far less of the project in mind at once.
- Plus electricity, plus upkeep, forever.
Or pay nothing extra
- A hosted model scores 82.7 for a few hundred rupees a month.
- The free option scores higher still, at zero.
- 2 lakh is 17 years of the subscription we pay today.
One good reason remains: nothing ever leaves the machine. That is a privacy decision, not a money one.
Hardware requirements and speeds from model cards; rental rates from RunPod, 12 Aug 2026.
What lasts
Free things vanish.
Marketplaces don't.
| Option | Outlook | Why |
|---|
| Paying a marketplace that routes to many providers | Most durable | Not a subsidy, so there is nothing to withdraw. If one supplier dies, it routes around it. |
| Paying per use, directly | Durable | You are a customer, not a cost centre. |
| A paid subscription | Repricing risk | One raised its entry price six fold this year. |
| A free tier from a big company | Fragile | Undefined limits, changeable at will, and five have already closed this year. |
So the plan should never depend on any single free tier staying alive.
The verdict
Change the model,
not the tool.
Do this first
Keep what we have, switch the model
880 per month, unchanged
Same subscription, one dropdown. Roughly 40 times more work from the same allowance, at about 90 percent of the best quality available anywhere. Costs nothing and risks nothing.
Then this
Claim the Jio offer
0 for 18 months
If it is the full paid plan, it unlocks Google's tool refilling every five hours with Claude's best model included. Free, already yours, and needs only a spare Google account to be safe.
Insurance
One payment to a marketplace
880 once, never again
Permanently raises a free allowance to 1,000 requests a day and lets us refuse any provider that trains on our code. The one thing here that cannot be taken away.
Total change to the monthly bill: nothing. Total gain: more headroom, a better ceiling, and a fallback.
Equally important
What we decided not to do.
- Do not buy hardware. Two lakh rupees buys a worse assistant than the free one.
- Do not switch to a cheaper subscription. Every one checked is a step down in quality, not a saving.
- Do not put the main Google account into Antigravity. A ban would take the app store account with it.
- Do not build on any single free tier. Five closed this year, two of them while this research was being written.
- Do not use MiniMax weights in the product without their name displayed. The licence forbids it.
The cheapest decision available was to change nothing about the bill.
Being straight about it
What we could not confirm.
?
Google's actual free limit
They publish no number anywhere. It is measured in work done, not requests. The only way to know is to use it and watch.
?
Whether the Jio plan is the full one
Its announcement lists older models and less storage than the current paid plan. Claim it and check the settings screen.
?
Several published prices
Some pricing pages load their numbers in a way that cannot be read automatically. Those are marked unverified here rather than guessed.
Every gap is listed in full in the research repository, not hidden.
Fidwise research
Better tools,
same bill.
Three research sweeps, more than 40 options, every price and privacy clause traced to the company that published it.
Compiled 12 August 2026. Prices and free tiers move quickly, so treat every figure as true on that date and no later.